Self-learning positioning using inferred context ... - SatNav Model Page

ICT–248894 WHERE2

D2.6

ICT–248894 WHERE2 D2.6 Final: Self-learning positioning using inferred context information Contractual Date of Delivery to the CEC: M34 Actual Date of Delivery to the CEC: 25.07.2013 Editor: Igor Arambasic Authors: Igor Arambasic, Javier Casajus, Tatiana K. Madsen, Jimmy J. Nielsen, Mathieu des Noes, Benoît Denis, Marios Raspopoulos, Julien Stéphan, Mohamed Laaraiedh, Bernard Uguen Participants: AAU, CEA, SIG, SIR, UPM, UR1 Work package: WP2 – Heterogeneous Context-aware Cooperative Positioning Est. person months: Security: PU Nature: R Version: 1.0 Total number of pages: 146 Abstract: An overview of research carried out inside WHERE2 T2.3, with respect to anchor-less indoor self-localization, tracking and movement prediction based on wireless communication properties is presented. We explain how to fairly assess the performance of different algorithms inside a realistic evaluation platform. The platform comprises a selection of typical indoor environments with refined layout descriptions, Ultra Wide Band (UWB) and Narrow Band (NB) measurements together with site-specific ray-tracing simulations (UWB, WB and NB) are available. Two main calibration free approaches are put forward. The first aims at retrieving the shape and the physical properties of indoor environments from transmitted radio signals, while performing anchor-less self-localization. The approach covers different algorithm designs and hardware implementation such as UWB and NB with spatial CIR discrimination. The UWB approach relies on multipath profiles, while spatial discrimination in NB approach is mainly through antenna sectorization. The precision of NB estimations is improved by introducing the confidence index concept in calculations and by proposing a novel mobility model designed to efficiently explore the unknown scenario. Additionally, investigations on the construction of environment maps based on indoor-to-outdoor environment are also conducted. The second group of solutions concerns mobility learning based on Hidden Markov Model (HMM), enhancing the tracking functionality.

1 / 146

ICT–248894 WHERE2

D2.6

Keyword list: Anchor-Less Localization, Context-Awareness, Directional Viterbi Algorithm, Hidden Markov Model, Indoor Mapping, Indoor-to-Indoor, Mobility Patterns, Narrow Band, Outdoor-to-Indoor, Ray-Tracing Simulations, Received Multipath Profiles, Reverberation, Sectorized Antennas, Self-Learning, Self-to-Self Channel Impulse Response, Simultaneous Localization and Mapping, Tracking Filters, Ultra Wide Band Disclaimer:

2 / 146

ICT–248894 WHERE2

D2.6

E XECUTIVE S UMMARY This deliverable provides the results carried out inside WHERE2 T2.3, with respect to mobility awareness, self-localization and environment characterization based on wireless communication networks. The outline of the task is found in appendix A.1. The document is structured in two parts, a body part providing a summary of reported activities organized topic-wise, and an appendix containing all related publications and reports produced within T2.3. The appendix part is intended to provide more detailed information on the conducted research and implemented methods. The body part of the deliverable summarizes in Section 2 the definition of a realistic evaluation framework which was chosen in order to comply with the generic professional context, (and to some extent, with the private context), as defined in Section 2.6 of D1.1b (pp. 15) [1]. It is indoors, with walls and obstructions, which restrict users’ movement patterns. The area covered by location service is less than 500 square meters, the mobility model is pedestrian, user independent, with maximum speed of 10m/s. Furthermore, this environment corresponds to that of previous WHERE1 measurement campaigns [2], hence the real measurements are complemented by deterministic ray-tracing simulations of impulse responses. Additionally, since the extraction of context-aware features is not constrained only to indoor-to-indoor scenarios, definitions of both: indoor-to-indoor and indoorto-outdoor synthetic environments have been conducted. Simulations are calculated by three partners of WHERE2 who elaborated their own ray-based solution namely PyRay, Volcano and 3DTruEM used for Ultra Wide Band(UWB), Wide Band (WB) and Narrow Band (NB) simulations respectively. Various location-dependent metrics can then be extracted from the obtained channel impulse responses (CIR) to feed indoor mapping and anchors-less localization algorithms. A refined layout description is also provided as it is of interest for mobility pattern learning or map-aided localization. A technical report concerning these ray-tracing tools together with the description of the synthetic environment, can be found in A.2, A.3 and A.4 . The application of ray tracing tools for indoor localization applications are investigated and validated in A.5. Furthermore, in order to allow efficient dissemination and exploitation of T2.3 synthetic environment, a Microsoft Sql Server (MSSQL) scheme format, identical to WP4 database, has been implemented as explained in Appendix A.7. Section 3 provides an overview of investigations with respect to self-localization and environment characterization based on wireless communication networks. Two novel approaches for indoor mapping and anchor-less localization are proposed. The first one is based on cooperative peer-to-peer UWB links, which exploit multipath propagation delays between various pairs of devices, while the second one applies radar approach together with spatial discrimination of CIRs implemented on NB antennas (100MHz bandwidth). Both algorithms are tested inside the synthetic ray-tracing environment and the obtained results are presented in detail in Appendices A.8, A.9, A.10, A.11, A.12 and A.13. Furthermore, the extraction of context-aware features from indoor-to-outdoor scenario, which aims at complementing the construction of indoor-to-indoor environment maps, is also conducted. The approach characterizes the radio wave propagation, when one terminal is inside a building (e.g. WiFi AP) and another one is moving along an outdoor trajectory in the close vicinity. The results presented in appendix A.14 demonstrate that coarse information about the building (e.g. building shape, presence of windows) may be extracted. New solutions for learning the users’ mobility patterns, in order to enhance the localization and tracking functionalities, are developed and described in Section 4. Here, we are trying to estimate the coordinates of a whole trajectory and not just of a single point, thus knowing mobility habits conditioned on the current environment can help obtaining a better estimation. E.g., smoothening of data by applying averaging filter can be done if a trajectory is continuous and is unlikely to make sharp turns. Hence, the focus is put on learning

2 / 146

ICT–248894 WHERE2

D2.6

mobility patterns through insite Hidden Markov Models (HMM) to assist tracking with the knowledge of the most probable routes followed by a user. The proposed mobility selflearning technique and statistics about the users’ behaviours (online/offline) are expected to assist either localization and tracking algorithms (i.e. noticeably those developed in T2.1, e.g. the TOA-based Maximum Likelihood location estimator used herein) or even fingerprinting based techniques (e.g. those developed in T2.2, including map-aided versions). More information with respect to this novel method can be found in appendices A.15 and A.16. Based on the work in D2.3 [3], the synthetic test environment has evolved from 3 independent MATLAB based files to a common WHERE2 SQL database which includes all the RT simulation data. In order to achieve this goal a MATLAB routine has been developed and implemented to populate the database. Indoor-to-outdoor simulations have been completed while new narrowband RT simulations of a single room based on a denser grid have been produced, supporting the work on environment learning. New UWB simulations have also been performed (in the same synthetic environment) to validate Hidden Markov Model based algorithm proposed to learn user movement. Additionally, UWB simulations of the synthetic environment have been post-processed in order to extract location dependent parameters (i.e. RSS and TOA). The simulations based location-dependent parameters (LDPs) have been compared to those extracted from measurements. This comparison pointed out that ray tracing UWB simulations can fairly estimate these LDPs and hence complement/replace measurement fingerprinting database once a proper description of propagation environments and antennas is performed. The algorithm for anchor-less self-localization based on CIR spatial discrimination of narrowband signals is expanded to offer the solution to the general SLAM problem. The method is improved by introducing the confidence index concept in calculations and by proposing a novel mobility model designed to efficiently explore the unknown scenario. The initially proposed self-localization algorithm based on cooperative peer-to-peer UWB links has been implemented with more realistic UWB channel impulse responses inside a common synthetic indoor environment. These responses have been pre-processed in order to extract a larger set of multipath delays (to combat small-scale fading), considering realistic Rx constraints, and also to adapt to the sampling rate of the algorithm. The novel part includes the fusion of local information in groups of 3 nodes for more robustness. Afterwards, the full SLAM estimation algorithm has been evaluated in different spatial configurations. The concept of extraction of context-aware features and localization from indoor-to-outdoor data has been briefly introduced in D2.3. Since then, the indoor/outdoor synthetic environment has been completed, using notably the final version of the highly realistic indoor-outdoor channel model elaborated in WP1 [4]. Some of the obtained indoor/outdoor channel realizations have been then characterized and analyzed in order to demonstrate the feasibility and interest of the proposed technique. With respect to Hidden Markov Models for mobility learning and tracking enhancement, the improved approach uses a slightly different abstraction model, which allows to directly encode also (fast) user movements, where the user jumps between non-neighbour grid points. Further, and most importantly, a realistic evaluation scenario based on the synthetic environment described in Section 2 is used to evaluate the proposed algorithms. When compared to D2.3, one line of research, namely, Mobility Learning Enhanced WiFi RSSI Fingerprinting Localization, is not present here. This contribution is linked directly to T2.2 location system (based on WiFi RSSI fingerprint), which is enhanced by incorporating historical information from previous movements inside the building, and hence was shifted to the T2.2. The motivation for this modification was to focus the task T2.3 on map making and indirect use of the map (whose constraints impact the probability of mobile patterns), while T2.2 would handle the comparison of fingerprinting approaches and the direct map-

3 / 146

ICT–248894 WHERE2

D2.6

aided positioning (i.e. setting geometrical constraints to the estimated locations). Besides the already mentioned publications which have strengthened the partners’ cooperation inside the task, the cross-work package collaboration should also be emphasized. The synthetic environment, based on RT simulations, was successfully used in T3.2 for evaluation of location-based relaying policies as can be seen inside Appendix A.6. The investigations of indoor-to-outdoor context information, presented in A.14, use realistic indoor-outdoor channel model elaborated in WP1 and described in D1.8 [4] . Furthermore, in cooperation with WP4, the RT simulations database format is modified and will eventually be made publicly available together with WP4 measurement data as one WHERE2 database.

4 / 146

ICT–248894 WHERE2

TABLE

OF

D2.6

C ONTENTS

1

Introduction

9

2

Synthetic test environment 2.1 Geometric Description of the Synthetic Environment . . . . . . . . . . . . 2.2 Applications of the Proposed Synthetic Environments . . . . . . . . . . . . 2.3 Delivered Database of the Synthetic Environment . . . . . . . . . . . . . .

3

Anchor-less opportunistic indoor terminal localization 16 3.1 Anchor-less Self-Positioning Based on Sectorized Narrowband Antennas . 16 3.2 Anchor-less Self-tracking and Room Dimensioning Based on Cooperative UWB Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 3.3 Extraction of Context-Aware Features and Localization from Indoor-to-Outdoor data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

4

Tracking and movement prediction 4.1 Hidden Markov Model based Mobility Learning . . . . . . . . . . . . . . . 4.1.1 Fundamental Study of HMM based Trajectory Filtering . . . . . . . 4.1.2 Modified HMM Based Algorithms and Realistic Case Study . . . .

27 27 27 29

5

Conclusions

31

References

11 11 11 14

34

A Appendix 35 A.1 Context-Awareness and Self-Localization in Wireless Networks: The WHERE2 Proposals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 A.2 Technical Report on Simulations of the Synthetic Environment . . . . . . . 45 A.3 PyLayers: An Open Source Dynamic Simulator for Indoor Propagation and Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 A.4 Efficient Ray Tracing Tool for UWB Propagation and Localization Modeling 64 A.5 Ray Tracing-Based Radio Propagation Modeling for Indoor Localization Purposes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 A.6 Model-based Evaluation of Location-based Relaying Policies in a Realistic Mobile Indoor Scenario . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76 A.7 Common Scenario Simulation Storage . . . . . . . . . . . . . . . . . . . . 82 A.8 Rectangular Room Dimensions Estimation Using Narrowband Signal and Sectorized Antennas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 A.9 Self-Positioning and mapping of rectangular rooms with sectorized narrowband antennas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 A.10 Anchor-less Self-Positioning in Rectangular Room Based on Sectorized Narrowband Antennas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 A.11 Joint Anchor-less Tracking and Room Dimensions Estimation through IRUWB Peer-to-peer Communications . . . . . . . . . . . . . . . . . . . . . 108 A.12 Benefits from Cooperation in Simultaneous Anchor-less Tracking and Room Mapping based on Impulse Radio - Ultra Wideband Devices . . . . . . . . 114 A.13 Cooperative Anchor-less Tracking and Room Mapping with Independent Computing Central Nodes and Application to Ultra Wideband Ray-Tracing Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 A.14 Extraction of Context-Aware Features and Localization from Indoor-to-Outdoor Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

5 / 146

ICT–248894 WHERE2

D2.6

A.15 Hidden Markov Model based Mobility Learning for Improving Indoor Tracking of Mobile Users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134 A.16 Directional Hidden Markov Model for Indoor Tracking of Mobile Users and Realistic Case Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

6 / 146

ICT–248894 WHERE2

D2.6

Authors Partner AAU

Name Tatiana K. Madsen

Phone/e-mail Phone: +45 99 40 86 32 e-mail: [email protected] Phone: +45 99 40 98 67 e-mail: [email protected] Phone: +33 4 38 78 97 82 e-mail: [email protected] Phone: +33 4 38 78 18 21 e-mail: [email protected] Phone: +357 22 32 52 40 e-mail: [email protected] Phone: +33 2 23 48 05 64 e-mail: [email protected] Phone: +34 91 54 95 700 ext.4006 e-mail: [email protected] Phone: +34 91 54 95 700 ext.4006 e-mail: [email protected] Phone: +33 2 23 23 50 75 e-mail: [email protected] Phone: +33 2 23 23 60 33 e-mail: [email protected]

Jimmy Jessen Nielsen CEA

Mathieu des Noes Benoît Denis

SIG

Marios Raspopoulos

SIR

Julien Stéphan

UPM

Igor Arambasic

Javier Casajus

UR1

Mohamed Laaraiedh

Bernard Uguen

7 / 146

ICT–248894 WHERE2

D2.6

List of Acronyms and Abbreviations 3GPP AP CDF CIR EKF FDP GLOS GNSS GPS HMM IR LDP LDR LOS NB NN NLOS PDP RMS RMSE RSSI RT RTT Rx SISO SLAM TB TOA Tx UWB WHERE2 WB WP

3rd Generation Partnership Project Access Point Cumulative Density Function Channel Impulse Response Extended Kalman Filter First Detectable Path Geometrical Line-of-Sight Global Navigation Satellite Systems Global Positioning System Hidden Markov Model Impulse Radio Location-Dependent Parameters Low Data Rate Line-of-Sight Narrowband Nearest Neighbor Non Line-of-Sight Power Delay Profile Root Mean Square Root Mean Square Error Received Signal Strength Index Ray Tracing Round Trip Time Transmitter Single Input Single Output Simultaneous Localization And Mapping Time Based Time Of Arrival Transmitter Ultra-wideband Wireless Hybrid Enhanced Mobile Radio Estimators - Phase 2 Wideband Work Package

8 / 146

ICT–248894 WHERE2

1

D2.6

I NTRODUCTION

Context-awareness is clearly seen today as a major stake and enabling feature of future wireless networks. The so-called context is usually understood in a wide sense, which may include physical elements of geometry, connectivity conditions or mobility aspects. Accordingly, the new concept of wireless context-awareness encompasses various functionalities like: • Environment characterization or mapping (i.e. the capability to retrieve a room shape, room volume, a building layout or even the nature of the wall materials around) • Anchor-less self-localization and tracking (i.e. the capability to retrieve the locations of mobile terminals in their nearby environment or with respect to other mobile fellows, with no fixed references) • Mobility self-learning (i.e. the capability to gradually learn and predict movement patterns and activity habits out of estimated trajectories). Ideally, these functionalities will be ensured out of standard radio communications and will enable a variety of unprecedented applications such as: • Context-based commercial services (e.g. location-based advertisements or offers in shopping centers, group/location-specific information broadcast) • Optimal network adaptation, i.e. depending on the local deployment conditions or operating environments (e.g. location-based interference mitigation or cancelation, packet relaying within identified clusters of cooperative users or based on the link obstruction information, etc.) • Assisted mobility or enhanced personal navigation in unknown environments (e.g. after detecting obstacles) • Etc. These new capabilities and applications are the catalyzer of the research activities inside task T2.3 where context-aware and self-learning methods are developed by using context information from measurements of physical properties (e.g. location-dependent radio metrics and/or Channel Impulse Response (CIR), etc.). This document, which forms the final report on the investigations inside task T2.3, is structured in two parts: a body part that provides a summary of the reported activities organized topic-wise, and an appendix containing all related publications and technical reports. The latter documents are meant to provide the readers with more detailed information on the activities if needed or wanted. Section 2: Synthetic test environment In order to allow T2.3 partners to realistically evaluate their self-learning positioning algorithms, a synthetic environment is defined based on a former UWB indoor measurement campaign carried out within WHERE1 project. Initially, the new synthetic environment provides channel impulse response (CIR) simulations of a 1m×1m 2D grid of 51 transmitters × 363 receivers spread inside a common indoor office environment. The simulations are performed using three ray tracing tools for three different bandwidths: PyLayers for UWB, Volcano for wide bandwidth and 3DTruEM for narrow bandwidth. The environment has been extended to cover indoor-to-outdoor scenarios in order to investigate extraction of context-aware features from indoor-tooutdoor data. Furthermore, a refined grid of points in one room is defined and simulated using the narrow-band 3DTruEM simulator in order to support the development

9 / 146

ICT–248894 WHERE2

D2.6

of self-localization based on sectorized narrowband antennas. Synthetic environment simulations have been exploited to model RSS and TOA and to complement fingerprinting database obtained from measurements. Realistic mobility patterns are generated using PyLayers simulator in the synthetic environment with the aim of evaluating the hidden Markov Model (HMM) based technique for estimating user’s movement. Finally, in order to deliver databases in the same format as WP4 databases, a Microsoft Sql Server (MSSQL) scheme has been defined to save all the CIR and the different simulation parameters using a common WHERE2 format. Section 3: Anchor-less opportunistic indoor terminal localization Two novel approaches for anchor-less self-localization inside the rectangular room are presented. The first algorithm enables simultaneous room mapping and selflocalization through cooperative peer-to-peer UWB links. It exploits multipath propagation delays observed over radio links between various pairs of devices to produce the required estimates. The second approach is based on eight 45 degrees sectorizedantennas with 100MHz bandwidth, which achieves the estimates utilizing radar type technique and self-to-self CIR. The approach is based on sectorized antennas as this kind of data is offered by 3DTruEM simulator, but can be implemented in any system capable of spatial discrimination of CIRs data. The results based on both synthetic data and deterministic ray-tracing simulations, are presented. Furthermore, we extracted context-aware features from indoor-to-outdoor scenario. The approach characterizes the radio wave propagation, when one terminal is inside a building (e.g. WiFi AP) and another one is moving along an outdoor trajectory in the close vicinity. Ultimately, the possibility to perform anchor-less indoor terminal localization using data obtained at the outdoor moving terminal is analyzed and discussed. Section 4: Tracking and movement prediction The movement habits of mobile users are exploited to improve their localization and tracking. This work assumes that users’ previous movement trajectories are stored in a database and uses the Hidden Markov Model framework to exploit this historical information for improving the estimation of the users’ movement trajectories. The proposed technique and statistics about the users’ behaviours are expected to assist either localization and tracking algorithms (i.e. noticeably those developed in T2.1) or even fingerprinting based techniques (e.g. those developed in T2.2). The novelty in this work lies especially in the fact that the used Hidden Markov Model encodes the movement direction in its state space, allowing to distinguish between movements from different directions, e.g., in a corridor. This idea is tested both in simple verification scenarios and in a realistic test scenario based on the synthetic test scenario presented in section 2. Section 5: Conclusions. References : Technical references are explicitly cited inside each annex contribution. However, more general references can be found in this section. Appendix A: The appendix contains a collection of already published or soon to be published articles or reports produced within task T2.3 of WHERE2.

10 / 146

ICT–248894 WHERE2

D2.6

2

S YNTHETIC TEST ENVIRONMENT

2.1

Geometric Description of the Synthetic Environment

The synthetic environment was defined during the first year of the WHERE2 project in order to allow T2.3 partners to realistically evaluate their self-learning positioning techniques. The proposed synthetic environment, presented in D2.3 [3], was initially based on the M1 measurement campaign carried out within WHERE1 project [2]. This measurement campaign consisted in measuring channel impulse responses (CIR) of UWB standard for radio links between a mobile Tx (302 positions) and 4 fixed receivers (302×4 = 1208 measurements in total). In the WHERE2 project, partners within the task T2.3 expressed the need for denser CIR realizations in different bandwidths to address heterogeneous and dynamic positioning techniques. Moreover, heterogeneous scenarios, which fuse short-range networks (UWB, WIFI, etc.) and long-range (Cellular) defined in D1.1b deliverable [1], require the definition and the simulation of an indoor-to-outdoor synthetic test environment. Ray-tracing based techniques appeared thereby as a valuable solution to complement real measurements by deterministic simulations of impulse responses as explained in [5], Appendix A.2 and A.5. Simulations were thus calculated inside the WHERE2 project based on ray-tracing solutions: • PyLayers to obtain ultra-wideband (i.e. > 500 MHz) simulations (see Appendixes A.3 and A.4) • Volcano to obtain wideband (i.e. ∼ 500 MHz) simulations • 3DTruEM to obtain narrowband (i.e. < 100 MHz) simulations The resulting data consist of CIR simulated over a 1m×1m 2D grid of 51 transmitters × 363 receivers spread on a common indoor synthetic environment that was parametrized to resemble the SIRADEL premises as illustrated in Figure 1. The points represent the grid chosen for performing simulations: the red points represent the transmitters while the blue points represent the receivers. The different components of the environments (i.g. doors, walls, etc.) are represented in different colours to highlight the difference between materials. In order to perform simulation with ray tracing tools, a detailed 3D site geometry, material properties together with a precise description of furniture is made available. This additional data is valuable for the generation and posterior emulation of plausible mobility patterns (and hence mobile radio traces) in the same environment. The parameters used for the synthetic environment by the different solutions are listed in Table 1. For each Tx-Rx pair defined in Figure 1, a CIR is obtained for each bandwidth (i.e. UWB, WB, and NB) as shown in Figure 2. 2.2

Applications of the Proposed Synthetic Environments

In this section we provide an overview of how the synthetic environment is used within WHERE2 to evaluate self learning localization techniques. The synthetic environment has been extended by SIRADEL to cover indoor-to-outdoor scenarios in order to investigate the extraction of context-aware features from indoor-to-outdoor data. A refined grid of points in one room is defined and simulated using the narrow-band 3DTruEM simulator with the aim of estimating rectangular room dimensions. In addition, synthetic environment simulations have been exploited to model RSS and TOA and to complement fingerprinting database obtained from measurements. Finally, realistic mobility patterns are generated using PyLayers in the synthetic environment in order to evaluate hidden Markov model based technique for estimating user’s movement.

11 / 146

ICT–248894 WHERE2

D2.6

Figure 1: 3D description of the synthetic environment (Dimension: 40m×16m). The red points represent the transmitters while the blue ones are the receivers. The different components of the environments (i.g. doors, walls, etc.) are represented in different colours to highlight the difference between materials.

Parameters Environment representation Transmitted power Maximum delay

Volcano 2.5D Digital Building Model (DBM) of the SIRADEL building with furniture 10 dBm

3DTruEM PyLayers Full 3D model

-45 dBm 300 ns

Isotropic antennas for Tx/Rx Isotropic antenna for Tx and sector antenna for Rx with theoretical radiation pattern: Antenna type 45 ◦ aperture in the horizontal plane Omnidirectional radiation pattern in the vertical plane Different azimuth: [0 ◦ , 45 ◦ , 90 ◦ , 135 ◦ , 180 ◦ , 225 ◦ , 270 ◦ , 315 ◦ ] Receiver -110 dBm any sensitivity Height of 1.5 m Transmitters Height of 1.5 m Receivers Table 1: Simulation parameters

12 / 146

ICT–248894 WHERE2

Narrow-band simulation between AP1 and MS38 (3DTruEM)

D2.6

Wide-band simulation between AP1 and MS50 (Volcano)

Ultra Wide-band simulation between AP1 and MS62 (PyRay)

Figure 2: Examples of obtained CIR for different bandwidths. Indoor-to-Outdoor Synthetic Environment This synthetic test environment is used with the highly realistic indoor-outdoor channel model elaborated in WP1 [4] to feed the investigations detailed in Section 3.3 and Appendix A.16 on extraction of context-aware features and localization from indoor-to-outdoor data. Channel realizations between three different WiFi access points (as indoor terminals) set in the same room in the first floor of the main building and a unique moving terminal outdoors are considered. Persistent clusters are extracted along each of these radio links and finely analyzed in two scenarios. First, we try to identify with which physical element of the map data they can be associated (window frames notably). Second, we assume that the access points (AP) locations are unknown and we try to estimate them by using location of the moving user terminal (a perfect location outdoors is considered). Simulations Concentrated on One Room with a Denser Tx-Rx Grid Narrow-band synthetic test environment is used for anchor-less self-positioning and simultaneous 2D estimation of rectangular room dimensions. The system presented in section 3.1 is based on narrow-band transceiver capable of simultaneous transmission and reception. The transmission is generated at omnidirectional antenna, while eight sectorized antennas with 45◦ apertures in horizontal are used for reception. In Appendix A.10, the presence of walls and corners is evaluated using round trip time (RTT) measurements based on spatially discriminated CIR of 8 sectorized antennas in the receiver, supported by this synthetic environment. Modeling of RSS and TOA Using Ray Tracing Simulations Beyond the initial interest for ray tracing deterministic simulations specifically and explicitly expressed in T2.3 (i.e. for SLAM and mobility learning, where the deterministic ap-

13 / 146

ICT–248894 WHERE2

D2.6

proach plays a critical role), it was rather obvious that these efforts made in T2.3 to ellaborate the synthetic environment could be mutualized afterwards for more generic purposes in T2.1/T2.2 and even WP3. The contribution presented in Appendix A.5, exploits the aforementioned three different ray tracing tools and the synthetic environment in order to investigate the importance of using such tools in localization techniques. These investigations have pointed out two important applications: substituting/complementing laborious measurement campaigns within fingerprinting based localization techniques and modeling of location dependent parameters (i.e. RSS and TOA). Investigations and measurements-simulations comparison showed that ray tracing tools can estimate these location dependent parameters once a proper description of propagation environments and antennas is performed. For both of these two applications, ray tracing tools represent a good alternative to measurement campaigns which are usually very laborious to carry out. Unlike measurements, the ray tracing tools offer also the possibility of updating the description of environments and antennas according to the studied scenario. Generation of Realistic Mobility Patterns In section 4 and Appendix A.15, a hidden Markov model based technique for estimating users’ movement given a sequence of inaccurate location coordinates is proposed. The approach is evaluated using realistic movement traces generated by PyLayers within the synthetic environment presented in subsection 2.1. The considered scenario is a situation where five agents are moving inside the indoor environment. The movement of agents is modeled using magnetic forces. The magnetic force model supposes that the agents and the layout environment can be assimilated to positive poles and the destination to a negative pole. Hence, at each time instant, each agent is under the influence of several magnetic forces: an attractive magnetic force which attracts the agent to its destination and several repulsive magnetic forces from the walls to avoid the collision. Each of those agents is able to measure four time of arrivals (TOAs) with respect to four anchors fixed at known positions in the indoor environment. These TOAs are estimated using the modified COST-231 multiwall model implemented in the network layer of PyLayers. A maximum-likelihood estimator implemented in the localization layer is then used to estimate agent locations from the four estimated TOAs. Additionally, this realistic indoor mobile scenario was successfully used in T3.2 for evaluation of location-based relaying policies as can be seen inside Appendix A.6. 2.3

Delivered Database of the Synthetic Environment

In order to deliver databases in the same format as the WP4 databases, a Microsoft SQL Server (MSSQL) scheme has been defined to save all the CIR and the different simulation parameters. This would allow to easily disseminate and exploit T2.3 results by WHERE2 partners and any other users. The adopted MSSQL scheme is shown in Figure 3 and detailed in Appendix A.7.

14 / 146

ICT–248894 WHERE2

D2.6

Figure 3: Database schema for Task 2.3 ray tracing simulations.

15 / 146

ICT–248894 WHERE2

D2.6

3

A NCHOR - LESS OPPORTUNISTIC INDOOR TERMINAL LOCALIZATION

3.1

Anchor-less Self-Positioning Based on Sectorized Narrowband Antennas

Anchor-less self-positioning approach based on the narrowband transceiver (100MHz), consists of one omnidirectional and 8 sectorized antennas. The transceiver supports full duplex technology, meaning it is able to transmit (omnidirectional antenna) and receive (sectorized antenna) simultaneously enabling self-to-self channel impulse response (CIR). By implementing self-to-self CIR on sectorized antenna a variety of applications are enabled and described in Appendices, including room volume determination [A.8], self-localization of two cooperating nodes [A.9] , and finally transceiver self-localization and rectangular room mapping, as described in [A.10].

Figure 4: Example of eight sectorized self-to-self CIRs When performing self-positioning we start with eight CIR observations (Figure 4), obtained at initial arbitrary position inside the rectangular room. Each CIR corresponds to one sector antenna pointing in the direction of the position of sub-image (i.e. the first column of 3 sub-images would correspond to upper left corner, left wall and down left corner CIR). By combining the information obtained at each receiver antenna the presence of walls and corners can be evaluated utilizing round trip time (RTT) measurements. Since the result of self-localization problem is provided in the form of distances with respect to the room walls, the construction of room 2D map is straightforward and is obtained simultaneously with localization parameter typical to the SLAM problem solution. When representing a rectangular room as a geometrical model, walls are seen as reflecting lines in a two-dimensional space, while corners are characterized by double reflections at two orthogonal walls. However, in real life, each received reflection is actually a sum of different reflection signals as a result of indoor- multipath propagation. These reflections are modeled with a single path since they cannot be discriminated due to receiver time resolution. This ambiguity is taken into account by making RTT reading dependent on transceiver’s bandwidth. As described in [A.8], each observation distance (d) in fact consist of two estimations, one corresponding to simple RTT readings and the other to distance ambiguity of the estimation: Dc E c d= f (RT T ), ( f (RT T ) + f (BW )) (1) 2 2 16 / 146

ICT–248894 WHERE2

D2.6

where c is speed of light, BW is the signal bandwidth. By restricting BW to 100MHz the precision of CIRs time scale is 10ns, which corresponds to distance resolution of 1.5m (0.5c/BW). In our environment this means that RTT measurement of 20ns, we cannot deduce that the wall is exactly at 3m distance, but we presume it is positioned between 3m and 4.5m from the transceiver. Naturally, if BW is larger the ambiguity of the measurement is decreased.

Figure 5: Estimation of walls and corners positions based on RTT of self-to-self CIRs obtained at sectorized Rx antenna. The arrows are pointing to the walls/corners. Their length depends on the BW ambiguity, and corresponds to two distance estimations in equation (1) Eventually every node position offers six estimations of each wall, four corresponding to corner estimations and two to direct wall positions. If differences in relative positions of the transceivers at every self-to-self CIR observation are known, wall estimation data can be successfully combined to produce adequate localization estimates. The geometrical description of two independent transceiver estimations is seen in Figure 5. The length of the arrows depends on the BW ambiguity, and corresponds to two distance estimations in equation (1). In order to produce the wall estimate, 12 observations of the corresponding wall are projected on the straight line as mathematically described in [A.8 and A.9]. We evaluated the approach on narrowband synthetic environment with the emphasis on a test room generated with higher spatial resolution of 0.5×0.5m. The office is actually a reception hall with four doors (2 at upper left corner, one at upper right corner, one at the center of the bottom wall). Its furniture is composed of metallic bookcase (on the left wall) and two wooden tables. The room is not perfectly rectangular with approximate size of 5.1m×7.3m and 2.4m height. The cumulative density function (CDF) of relative errors obtained for dimension of X and Y are depicted in Figure 6 obtained with the approach developed in [A.8] by simply calculating the median of 12 estimations for each wall. Nevertheless, the CDFs show that in more than 80% of calculations the relative error of the obtained dimensions is below 15%, while in 50% of calculations the error would be less than 11%. These results are improved in [A.9] where two discriminations of inadequate estimations are introduced: 1) If the transceiver is in a corner its distance with respect to each of corner walls cannot be smaller than 0.25m and larger than 1.5m. We assume that the transceiver is

17 / 146

ICT–248894 WHERE2

D2.6

Figure 6: CDF of relative estimation error of room dimensions (east-to-west wall denoted as X; north-to-south wall as Y) in a corner if two orthogonal readings, together with the corner observation between them, produce the maximum energy reading at 0ns. 2) If both transceivers are inside the room, walls cannot be in-between. If any of these constraints is not met, the distance estimation is regarded as unreliable and is not taken into account. The mean value of the rest of observations produces the wall estimation. The CDF of relative error obtained for horizontal and vertical 2D room dimension with this method is depicted in Figure 7. The CDF shows that in over 85% of calculations the relative error of the obtained dimensions is below 15%, while in 50% of calculations the relative error of room dimensions would be less than 7%. When compared to the results presented in A.8 the error slope is shifted to the left by approximately 5%. Additionally, the CDF confirms that the estimation of horizontal walls is susceptible to higher errors as furniture covers higher percentage of these walls In [A.10], we assume that a transceiver is a part of an “intelligent“mobile robot platform capable of moving, updating its positions, and keeping track of past positions together with the corresponding measurement data. As a result, a number of required measurements for reaching certain precision can be extracted. The novel part of self-localization algorithm is the incorporation of distance estimation confidence indicator (W), which is defined as the instantaneous received power of the CIRs strongest element (corresponding to RTT measurement): W = 10

max(CIR) 10

(2)

By taking into account the measure confidence indicator, the distance estimate of wall m is seen as: m Wi,n (3) dm = ∑ W m di,nm i=1:6,n=1:N 0 where m,i,n describe m-th wall and i-th estimation of n-th observations of the total number of N observations (i=1..6, m=1..4, n=1..N). The W0m is the sum of all confidence indicators for one wall and is used as an indicator of the wall estimation precision. The CDF of relative estimation error of room wall dimensions is given in Figure 8. This relative error is directly related to the positioning error since transceiver position is provided

18 / 146

ICT–248894 WHERE2

D2.6

Figure 7: CDF of relative estimation error of room dimensions (east-to-west wall denoted as X; north-to-south wall as Y

Figure 8: CDF of wall dimension estimation relative error as a function of number of observations

19 / 146

ICT–248894 WHERE2

D2.6

as estimated distance with respect to the four room walls. It can be observed that CDF curve has two slopes: the greater one for CDF value of up to 0.8, 0.9 and the lower one for larger values. This property can be observed for all curves indicating the existence of precision limit of the approach regardless of the number of observations. The existence of this limit is easily noticed if we consider a transceiver located closely to the wall (less than 0.5m). In this case, the reflection originated at the wall will be of large strength and its peak will be inside the first time sample. According to the eq. (1) we would have two estimations, pointing to 0m and 1.5m. Since the weight of these estimations is very high the final wall estimation would be close to 0.75m, which compared to true distance value would suppose relative error of almost 50%. Nevertheless, the gain of larger number of observation points is seen in the low-slope region, which gets smaller if the number of observations is increased. For example, if the trajectory is limited to 2 steps, the error is between 20% and 50%, while for 10 observations is between 10% and 25%. Besides this advantage, the relative error in 90% of cases is reduced approximately to the half, from 25% to 12%, if the number of observation points is incremented from 2 to 10.

Figure 9: CDF of average absolute positioning error as a function of number of observations The CDF distribution of the average absolute positioning errors is shown in Figure 9. The average positioning error is defined as the average error of 4 walls distances estimations. This result shows that the average error in the worst estimation scenario, in meters, is reduced from 1.3m to 0.85m when using 10 instead of 2 observations. On the other hand, it can be observed that there is close to 50% of simulated scenarios in which incrementing the number of observation points, the localization precision is not improved, as the gain between 10 and 2 points is very low, from 0.6m to 0.5m. Again, this is due to bandwidth limitations and due to number of observations close to the walls that produce erroneous estimations with over-confidence. Regardless of these limitations, the localization performance is improved with higher number of observation points.

20 / 146

ICT–248894 WHERE2

D2.6

The novel component includes the introduction of instantaneous received power of the CIR’s strongest sample as indicator of position estimate quality. Hence wall distance estimate is not found based on simple mean average but on the weighed average. Furthermore, a sum of all confidence indicators corresponding to the same wall is proposed for measurement of wall estimation precision. This parameter is used as fundamental part of the recommended mobility model developed with the aim of efficiently exploring the unknown environment. The algorithm is tested using synthetic data obtained with deterministic ray-tracing simulations of CIRs. It is shown that the even with 100MHz signal (up to 3m localization error per estimation due to sampling rate) the relative wall dimension estimation error is maintained in more than 80% of points below 9% and 12% for 10 and 5 observations respectively. The absolute positioning error with respect do real wall distance is below 0.6m in more than 70% of calculations obtained with 10 observations, while 0.8m error is obtained in the same percentage for only 2 observations. Besides confirming that this approach can be used for indoor localization and mapping, the analysis of the results affirms that better results are obtained with higher number of observations. In all contributions we consider four sectorized antennas to be perpendicular to the four walls and four antennas to point towards the corners. It is assumed that the selection of adequate antenna orientation is performed by physically rotating the mobile platform and selecting the position with minimum RTT readings or by using beam forming approach. Perhaps, it should be emphasized that the presented anchor-less self-positioning algorithm is also not restricted only to sectorized antennas. The essential part of this approach is spatial discrimination of the received signals which can also be achieved with 2D array of N omnidirectional antennas capable of beam forming. Following the same line, the algorithm is not restricted to 45◦ horizontal aperture either, and if this aperture is narrower the positioning results should be more accurate. Further work includes analysis of localization precision as a function of used bandwidth and horizontal aperture of sectorized antennas. Additional research to estimate number of walls in a non-rectangular room is also marked as valuable for the generic the SLAM solution. 3.2

Anchor-less Self-tracking and Room Dimensioning Based on Cooperative UWB Nodes

The Impulse Radio - Ultra Wideband (IR-UWB) technology is well known for benefitting from fine intrinsic multipath resolution capabilities, making it particularly attractive for localization-oriented applications in indoor -and more generally, confined- environments. In particular, among the multipath components resolved on the Rx side, echoes issued from simple sequences of electromagnetic interactions (e.g. single-bounce reflected paths) are often observed to be energetic, and hence, they tend to be more systematically detectable and consistent under mobility due to spatial correlation effects. The processing of their excess delays, which structurally convey information about the geometry of the involved radio links (e.g. see Fig. 10), thus makes possible the anchor-less estimation of the relative nodes’ coordinates {(xm , ym ), (xn , yn )} in the room, whose wall dimensions (Dx , Dy ) can also be determined simultaneously. According to a first approach proposed at the very beginning of WHERE2 (See Section A.11), one can exploit the estimated delays {τi }i=1..4 of the four strongest secondary paths, which are theoretically associated with the four single-bounce reflections in the room. The idea is to benefit from simple and explicit mathematical relationships, of the form τi = fi (xm , ym , xn , yn , Dx , Dy ). One more proposal, described in A.12, consists of operating with more than 2 cooperative nodes, contrarily to standard single-link approaches. The latter indeed suffer from inherent geometrical ambiguities due to e.g., combined Tx/Rx and

21 / 146

ICT–248894 WHERE2

D2.6

Figure 10: Single-bounce reflection model for 2 transceiver nodes (TRx) among 3, communicating over Line of Sight links in a rectangular room with arbitrary wall numbering (*) (left) and simplified received Channel Impulse Response accordingly (right). mirror flip uncertainties (See e.g. Section A.11). In a real system, the multipath profiles and delays observed between each pair of devices are practically estimated within low-complexity receivers, e.g. compliant with the IEEE 802.15.4a standard. Including the expected direct path, (at least) 5 delays {t j } j=0..4 measured at this receiver (e.g. sorted in descending energy order) still need to be mapped onto the a priori modeled {τi }i=0..4 , according to an arbitrary wall numbering. That is, one needs to associate each measured delay with a reflection onto the appropriate wall to determine the rule j ↔ i. Assuming systematic LOS channel configurations in the room, the first measured delay t0 can always be associated with the direct path and hence mapped onto τ0 . But 24 remaining mapping combinations must be tested for each peer-to-peer link. This data association problem can be solved either once for all during the initialization phase (i.e. considering the initially determined mapping can remain unchanged afterwards) or continuously updated in the steady-state tracking regime. Accordingly, at each time iteration under mobility, we assume i) the acquisition of the received multipath channel, ii) the estimation of the strongest multipath components, iii) the update of the data association (i.e. mapping between the indexes of detected and modeled paths) and iv) the continuous tracking of the estimated coordinates and dimensions through EKF filtering, fed with the resorted TOAs. In the initially proposed solution, a specific state initialization procedure based on the minimization of quadratic cost functions has been proposed and the Nearest Neighbor (NN) algorithm has also been optionally considered to continuously update the data association rule in the steady-state regime. The idea is to select the combination of Time Of Arrival (TOA) measurements that is the "nearest" to the value predicted by the tracking filter. This strategy is similar to an "inverse" innovation monitoring approach. Finally, to be even more robust with respect to poor filter state initialization, which can cause EKF filter instability (even in case of relevant data association), a cooperative tracking filter structure involving at least three cooperative nodes has been introduced, whose state vector contains the coordinates of all the involved devices, along with the room dimensions Dx and Dy . As a first step of our investigation, the performance of the previous algorithm was preliminarily assessed with synthetic TOA measurements, assuming that the estimated multipath delays are only affected by centered Gaussian noise. Results obtained with the proposed optimization-based scheme have been compared with a perfect state initialization, averaging over mobile trajectories simulated in a rectangular room of Dx = 10m and Dy = 20m under moderate pedestrian speeds. These preliminary results are found in Appendix A.12. However, considering groups of 3 nodes as a good compromise between low (combina-

22 / 146

ICT–248894 WHERE2

D2.6

torial) complexity and performance improvements through cooperation, only one node was selected so far as "central" and the 2 communication links with respect to the 2 other nodes were exploited to estimate simultaneously the room dimensions and the nodes’ locations. During the second half of the project, this concept has been generalized. Accordingly, each node of the group is successively considered as central, and the state initialization procedure described above is applied for each central node. The configuration that gives the minimum cost function is thus selected, and the corresponding initial state and configuration feed a unique tracking filter afterwards. Under the same simulation set-up as previously, the 50% percentile of the errors regarding estimated locations and room dimensions (i.e. median estimation errors) obtained with this circular cooperative strategy are presented in Tables 2 and 3 respectively, for both static and mobile devices and moderate speeds (< 0.5 − 1m/s). Again, in spite of a purely blind initialization and reasonably conservative TOA precision levels here (compatible with the time resolution usually available at real UWB receivers, e.g. 4ns in our case), anchor-less location performance appears to be on the order of that achieved by conventional anchors-based solutions, with a median error better than 1m in all the tested cases. Despite its simplicity and the use of exactly the same number of nodes, the "multiple central nodes" approach makes initialization more robust, and hence improves the final performance. A more detailed description of the new proposal, as well as additional simulation results, can be found in the Appendix A.13). Note that other tight fusion strategies between the three cooperative nodes could also be applied, but at the price of higher computational complexity (e.g. fusing or combining afterwards the outputs of EKF running independently at the 3 different central nodes).

Figure 11: Combination of circular computation "centers" in cooperative groups of 3 nodes.

Perfect State Init. Optimization-based State Init., 1 central nodes Optimization-based State Init., 3 central nodes

Static 0.62m 0.84m 0.82m

Mobile (Vmax < 0.5m.s−1 ) 0.86m 0.92m 0.89m

Table 2: 50% percentile of the location estimation error (i.e. ε50% in m) in a 5mx10m room

Perfect State Init. Optimization-based State Init., 1 central nodes Optimization-based State Init., 3 central nodes

Static 0.26m 0.65m 0.6m

Mobile (Vmax < 0.5m.s−1 ) 0.48m 0.65m 0.64m

Table 3: 50% percentile of the room dimensions estimation error in a 5mx10m room The proposed algorithms described above have been also evaluated considering more realistic received channel profiles and more realistic receiver constraints (assuming energy

23 / 146

ICT–248894 WHERE2

D2.6

detection with coarse timing resolution of 2ns). We have exploited IR-UWB ray-tracing simulations generated in the frame of T2.3 at Siradel premises (See Figure 12). More particulary, considering one Tx in a reference rectangular room and the CIRs generated over a uniform 1m× 1m grid of Rx positions in the same room, the 10 strongest peaks/bins in the received energy profiles over each link have been selected as potential candidates to feed our tracking and mapping algorithms. It appears that the CIRs of approximately 40 % of the tested Rx positions could not be exploited directly to solve the tracking and mapping problem, as formulated previously. In such configurations, at least one multipath corresponding to an expected reflection is indeed apparently missing at first sight, most likely due to overlapping paths under relaxed timing resolution, in particular in Rx locations close to the walls. Among the remaining 60-% of exploitable CIRs however, key illustrating configurations have been tested to practically assess the performance of our estimation algorithms (See A.13). The results tend to show that the multipath delays extraction procedure (out of the estimated CIR) could be improved further (e.g. applying local maxima detection scheme) and/or that the filtering algorithm must cope with missing or incomplete observations in certain locations along continuous trajectories. Indeed a simple algorithm as the one used in this study may not be robust enough in all the geometrical configurations. The analysis of viability of larger number of cooperating nodes should be made in future. This would include the influence of larger processing load and additional signalling messages in the network on the feedback of estimated channel impulse response to the processing unit. 3.3

Extraction of Context-Aware Features and Localization from Indoor-to-Outdoor data

Investigations on the construction of environment maps are usually limited either to the indoor or outdoor environment. However, using the technique to geo-locate the indoor terminals in a global environment (in which building? where inside the building?) might also be of a great interest. That is why we investigate an innovative technique that consists of the extraction of context-aware features from indoor-to-outdoor channel data. The main idea lies in carefully analyzing radio wave propagation when one terminal is inside a building and another terminal is moving outside in close vicinity of the building and extracting coarse information about the building (e.g. building shape, presence of windows). We study the feasibility of such a technique and search after the channel parameters that may provide the most relevant information. At this stage, investigations are based only on channel simulations. The technique is more precisely based on the detection of the static clusters that may for instance be generated by the window/door frames. The investigations make use of the original indoor-outdoor synthetic test environment presented in D2.3 [3], the highly realistic indoor-outdoor channel model elaborated in WP1 (see [4] for more details) and an algorithm to detect the presence of static clusters in successive (in space or time domain) channel responses described in D1.8 [4]. This cluster detection algorithm gathers multiple simulated ray paths based on angular and delay spreads criteria. As detailed in appendix A.14, the analysis has been done for three different Access Points (which act as indoor terminals) set in the same room in the first floor of a main building. We define a moving terminal outdoors that follows a 90 m-long linear trajectory in front of this building. Indoor-to-outdoor simulations have been computed at the frequency of 2462 MHz (i.e. channel 11 of the WiFi standard) between each Access Point (AP) and the moving outdoor terminal. An omnidirectional antenna is considered for all APs as well as for the moving terminal. It is set at the absolute height of 1.85 m for both. Persistent clusters have been then extracted from the successive radio links simulated for each Access Point (AP). We identified the persistent clusters having a constant angle prop-

24 / 146

ICT–248894 WHERE2

D2.6

Figure 12: Tested reference room in the T2.3 synthetic environment (SIR building) (a); Zoom on the reference room, showing favorable and harmful Rx positions in terms of overlaping paths (b) and examples of favorable (c-left) and harmful (c-right) CIRs, with expected delays from the actually reflected paths (Theory) and estimated delays from the strongest eenergy samples (Max).

25 / 146

ICT–248894 WHERE2

D2.6

erties (angle-of-departure or in angle-of-arrival) and we finely characterized those that are present over at least 5 successive intervals (see appendix A.14). We calculate the location of the interaction point or “bright spot” related to the angle-of-departure and angle-of-arrival mean properties of each cluster for each interval (i.e. location of the outdoor mobile terminal every 5m along the trajectory), considering that the main contribution(s) in high-power clusters are more likely to result from one single interaction, and identify with which physical element of the geo map data (window frames notably) it can be associated. Correlation with specific elements of the map data, either a window or a building edge, is obvious. We obtain a global mean error of 0.97 m only, and a standard deviation of 1.17 m. Lastly, we consider the scenario where the AP locations are unknown. We thus try to estimate them by assuming exact knowledge of the moving user terminal location (i.e. a perfect GPS location outdoors) and using the mean delay statistic, mean angle-of-departure statistics and mean bright spot location obtained for each extracted static cluster. As detailed in appendix A.14, the obtained estimation is good: the mean error on the AP location is globally 1.47 m. This result thus confirms the feasibility and interest of the proposed technique. Possible application may be to combine such a technique with wide band and Multiple Input Multiple Output antenna based outdoor terminals (typically LTE/LTE-A mobile phone or mobile sounders) equipped with GPS and able to scan outdoor/indoor networks (e.g. 4G plus WiFi) in order to localize the indoor APs in a global environment. Then the localized indoor APs can help in the global localization of the indoor terminals attached on them.

26 / 146

ICT–248894 WHERE2

4

T RACKING

D2.6

AND MOVEMENT PREDICTION

In many applications a point of interest might not be reduced only to a single location estimation of a device, but on a whole trajectory of a moving object. A moving object tracking refers to a process of establishing the spatial and temporal relationships between moving and stationary objects, in many cases between a moving device and a map. Since we are trying to estimate the coordinates of a whole trajectory and not just of a single point, knowing some properties of an object and the environment can help obtaining a better estimation. E.g., smoothening of data by applying averaging filter can be done if a trajectory is continuous and is unlikely to make sharp turns. In a similar way, knowledge about walls and obstacles can help to improve the accuracy of a trajectory estimation of a moving object. On the other hand, since the building layout also has a direct impact on the user’s movements (in the sense that the user is not able to walk through walls) historical movement traces of a user can be investigated and exploited to learn the user’s typical movement patterns and thereby improve tracking performance. The motivation behind this approach is based on the fact that if a turn was observed in many training trajectories, then a presence of an obstacle in the room can be deduced with high certainty. The proposed mobility self-learning technique and statistics about the users’ behaviours (online/offline) are expected to assist either localization and tracking algorithms (e.g. the TOA-based Maximum Likelihood location estimator developed in T2.1) or even fingerprinting based techniques (e.g. those developed in T2.2, including map-aided versions). The considered scenario is a grid-representation of an indoor building environment where the user moves between rooms through doorways and corridors, similar to the environment specified in section 6.1.5 in deliverable D1.1b [1]. 4.1

Hidden Markov Model based Mobility Learning

In this subtask, we have considered Hidden Markov Model (HMM) based filtering algorithms that use previous observations to estimate a user’s most likely movement trajectory, given a sequence of inaccurate location coordinates. The proposed, novel HMM-based algorithms are able to learn user habits by discriminating between different movement directions when populating the state transition probability matrix from training data. The proposed algorithms are intended for post-processing the location estimates of a traditional indoor localization system, see Figure 13. Radio measurements, e.g., RSS, ToA

Localization system

(x(t),y(t))

HMM model

(x*(t),y*(t))

Figure 13: The proposed HMM based algorithms are used to post-process the output of a localization system. The initial work within this subtask is summarized in section 4.1.1 and presented in detail in Appendix A.15. The second iteration of this work is summarized in section 4.1.2 and presented in detail in Appendix A.16. 4.1.1

Fundamental Study of HMM based Trajectory Filtering

The principle of the proposed algorithm is to first use training data to adjust transition probabilities in a Markov Chain representation (see Figure 14) of the considered building layout, thereby learning the mobility model. Afterwards, this mobility model is incorporated into

27 / 146

ICT–248894 WHERE2

D2.6

a filtering algorithm based on the Viterbi approach referred to as the Adaptive Viterbi algorithm, which takes the learned mobility behavior into account when estimating the user’s most likely trajectory.

Figure 14: Markov Chain representation of geographical area and mobility model. Transition probabilities are updated according to historical movement information. Both an off-line and on-line version of the algorithm have been developed differing in the way training data is collected and processed. In off-line mode, the training phase and construction of a matrix with transition probabilities is done before it is used for position estimation, whereas in online mode every output of the filtering algorithm is used to update the transition matrix. The conducted preliminary simulation results support the claim that the adaptive approach based on a self-learned mobility can improve the estimation accuracy. The parameter used to evaluate the performance is a mean error between a smoothened noisy measurement and a real node position. Table 4 compares different approaches: (1) simple moving average approach; (2) HMM-based filtering without mobility model knowledge (all directions are equiprobable); (3) HMM-based filtering assuming a straight movement of an object and (4) HMM-based filtering using self-learned mobility model. Filter (1) Moving average (2) Viterbi (3) Directive Viterbi (4) Adaptive Viterbi

Error off-line 1.12 0.67 0.67 0.56

Error on-line 1.12 1.13 0.99 0.78

Table 4: Comparison of different approaches. The unit is grid length. The HMM based approach is completely decoupled from a position estimation algorithm and can be used for post-processing position estimations and removing measurement noise. The proposed algorithm calculates the most probable path of a user by considering his past movement traces. For some implementations it can be a useful add-on feature that improves the performance – especially in cases when modification and fine-tuning of a localization algorithm itself is not possible. The presented approach will work with any localization algorithm.

28 / 146

ICT–248894 WHERE2

4.1.2

D2.6

Modified HMM Based Algorithms and Realistic Case Study

The second contribution within this subtask is a second iteration on the HMM based mobility learning algorithm. This second iteration improves the work presented above in two ways: 1) by using a realistic case study to test the algorithm, and 2) by allowing the transitions between non-neighbor grid points, which now detects easily the movements of users with varying movement speeds. On the other hand, this second iteration is currently not supported in the on-line mode. As before, the proposed algorithms are initially tested using simple artificially generated movement traces. However, secondly, in this iteration, they are evaluated using realistic movement traces generated with the PyLayers simulator [6]. PyLayers is an open source radio simulator based on a graph description. It is designed to simulate complete dynamic scenarios with realistic movement of users inside a building, the transmission channel estimation for multiple radio access technologies, and the position estimation of users. Evaluation results for the simple test scenario with two oppositely intersecting trajectories demonstrated a significant improvement of location accuracy with the Directional HMM algorithm, as shown in Figure 15. Empirical CDF

1 0.8

F(x)

0.6

0.4 Localization system Loc. Sys. + std. HMM Loc. Sys. + dir. HMM

0.2

0 0

0.5

1

1.5

2 2.5 RMSE [m]

3

3.5

4

4.5

Figure 15: CDF of RMSE for artifical test scenario. Further results for a scenario with realistic simulation based movement trajectories also showed improvements for 60% of the cases, however only if the HMM models are trained with (usually unknown) true trajectories. When trained with inaccurate location estimations, the HMM based algorithms showed no benefit compared to just using the localization system as seen in Figure 16. This much lower improvement for the realistic movement traces we believe is partly due to insufficient habitual behavior in the realistic movement traces, as waypoints are randomly and independently chosen for users. Therefore, it would be interesting to test the algorithms with realistic movement traces where user movements have stronger and more repetitive patterns. However, this is currently not yet supported by the PyLayers simulation tool.

29 / 146

ICT–248894 WHERE2

D2.6

Empirical CDF

1 0.8

F(x)

0.6

0.4 Localization system Loc. Sys. + std. HMM Loc. Sys. + dir. HMM

0.2

0 0

1

2

3 RMSE [m]

4

5

6

Figure 16: CDF of RMSE for realistic evaluation scenario where observation data is used for training.

30 / 146

ICT–248894 WHERE2

5

D2.6

C ONCLUSIONS

This deliverable summarizes the WHERE2 investigations on the topic of self-learning positioning using inferred context information. A special focus has been put forward on the generation of a synthetic environment (for test and performance evaluation purposes), indoor mapping, anchors-less localization and mobility learning. Three different ray-based simulators (UWB, WB, and NB) applied on a highly realistic indoor and indoor-to-outdoor synthetic environment allowed the generation of a very rich database that has been exploited within T2.3 to evaluate self-learning localization techniques. Investigations showed that the proposed simulations are reliable in modelling location dependent parameters (such as RSS and TOA) and in complementing fingerprinting measurement databases, which are usually very laborious to establish. The synthetic environment has been also exploited to provide inputs for techniques that aim to estimate rectangular room dimension or user movement within indoor environments. It should be emphasized that the synthetic environment was built taking into account only the requirements expressed by T2.3 partners. As such, some features (required afterwards by partners not involved in T2.3) like mobility and cooperation are found missing. Hence, in the second half of WHERE2, this synthetic environment was extended and was taken as basis for more generic scenarios that have been developed in WP1 in order to evaluate localisation techniques from entire WHERE2 project perspective. Interestingly the generated simulated data have also been harmonized and formatted following the same database SQL formalism as for WP4 measurement to ease further dissemination into the interested scientific community. Two novel approaches for indoor mapping and anchors-less localization are presented; one based on cooperative peer-to-peer UWB links which exploits multipath propagation delays between various pairs of devices, and another based on eight narrow band (100MHz), 45 degrees sectorized, antennas that utilize self-to-self CIRs to obtain the position. Their performance was assessed by means of empirical CDFs of the absolute RMSE and relative error (with respect to the real nominal values). The novel component of NB sectorized approach includes the introduction of instantaneous received power of the CIR’s strongest sample as indicator of position estimate quality. Hence wall distance estimate is not found based on simple mean average but on the weighed average. Furthermore, a sum of all confidence indicators corresponding to the same wall is proposed for measurement of wall estimation precision. This parameter is used as fundamental part of the recommended mobility model developed with the aim of efficiently exploring the unknown environment. The CDFs show that the relative wall dimension estimation error is maintained in more than 80% of points below 9% and 12% for 10 and 5 observations respectively. The absolute positioning error with respect do real wall distance is below 0.6m in more than 70% of calculations obtained with 10 observations, while 0.8m error is obtained in the same percentage for 2 observations. It is worth mentioning that even though the positioning error is reduced by increasing the number of observations, the existence of two slopes in all CDFs indicate the existence of precision limit of the approach. This is easily explained if we consider a transceiver located closely to the wall (less than 0.5m). In this case, the reflected component of received signal is very high, so its weight in the position estimation would be high and the wall prediction would be close to 0.75m, which compared to true distance value would suppose relative error of almost 50%. The cooperative UWB approach has also been evaluated with more realistic channel profiles obtained by means of UWB ray-tracing simulations. Groups of 3 nodes have been considered as a good compromise between low (combinatorial) complexity and performance improvements through cooperation, only one node was selected so far as "central" and the 2 communication links with respect to the 2 other nodes were exploited to estimate simultaneously the room dimensions and the nodes’ locations. During the second

31 / 146

ICT–248894 WHERE2

D2.6

half of the project, this concept has been generalized as each node of the group is successively considered as central, and the state initialization procedure is applied for each central node. The results show that in spite of a purely blind initialization and reasonably conservative TOA precision levels here (compatible with the time resolution usually available at real UWB receivers, e.g. 4ns in our case), anchor-less location performance appears to be on the order of that achieved by conventional anchors-based solutions, with a median error better than 1m in all the tested cases. Despite its simplicity and the use of exactly the same number of nodes, the "multiple central nodes" approach, tends to improve the final performance, mostly in the worst case error regime (formerly caused by poor initialization), while the mean error is only slightly improved. Method UWB UWB UWB NB sectorized NB sectorized NB sectorized

Room 5mx10m 5mx10m 5mx10m 5mx7m 5mx7m 5mx7m

Scenario Perfect State Init. Optimization-based State Init., 1 central nodes Optimization-based State Init., 3 central nodes 10 observations 5 observations 2 observations

Error 0.62m 0.84m 0.82m 0.52m 0.55m 0.65m

Table 5: 50% percentile of the location estimation error using WHERE2 self-localization approaches in 3 different rooms The 50% percentile of the errors regarding estimated locations obtained with UWB cooperative strategy, together with the results of NB sectorized approach, are presented in Table 5. The simple comparison of the obtained results might be misleading since these methods differ at hardware, algorithmic and even at conceptual level. The UWB approach includes at least three cooperative nodes while NB approach only uses a single node that is moving and obtaining the measurements at different locations. The TOA precision level of 4ns are very conservative taking into account that the NB approaches precision is of 10ns which might be the reason that the difference between the results is not so high. Additionally, the results obtained for NB approach correspond to the smaller room so its relative error would be larger. The analysis of the results affirms that both approaches can be used for indoor localization and mapping obtaining the results in line with conventional anchors-based solutions with a median error better than 1m. Investigations on the construction of environment maps are usually limited either to the indoor or outdoor environment. However using this kind of technique to geo-locate the indoor terminals in a global environment (in which building? where inside the building?) might also be of a great interest. This is why the feasibility of an innovative technique that consists in the extraction of context-aware features from indoor-to-outdoor channel is also studied. The main idea lies in finely analyzing radio wave propagation when one terminal is inside a building and another terminal is moving outside in close vicinity of the building and extracting coarse information about the building. Furthermore, the scenario where the unknown access points locations are estimated, by assuming exact knowledge of the moving user terminal location (i.e. a perfect GPS location outdoors), is analysed with the mean error of the access point location, i.e. 1.47 m. These results confirm the possibility of extracting the context-aware data from indoor or outdoor environment and its feasibility in practice. The contribution in Section 4, which concerns the exploitation of users’ habits to improve tracking of mobile users, successfully demonstrated that the proposed directional Hidden Markov Model (HMM) is able to distinguish between movements from different directions. This enables the use of HMM model to improve estimated movement trajectories from a localization system, by taking historical information about users’ typical movements

32 / 146

ICT–248894 WHERE2

Filter (1) Moving average (2) Viterbi (3) Directive Viterbi (4) Adaptive Viterbi

D2.6

Error off-line 1.12 0.67 0.67 0.56

Error on-line 1.12 1.13 0.99 0.78

Table 6: Tracking performances comparison of different approaches for learning the user’s movement. The unit is grid length. into account. The conducted simulation results seen in Table 6 support the claim that the adaptive approach based on a self-learned mobility can improve the estimation accuracy. The parameter used to evaluate the performance is a mean error between a smoothed noisy measurement and a real node position. A more realistic case study, based on movement traces generated using the PyLayers tool for a derivative of the synthetic environment, presented in Section 2, did not show any benefit of using the proposed model compared to using the initial estimation from the localization system. We believe that this is partly due to the generated movement traces, which do not contain strong enough habits as they result from movement way points being chosen completely randomly. Further studies of less random movement trajectories are therefore needed to fully understand the capabilities of the proposed models.

33 / 146

ICT–248894 WHERE2

D2.6

R EFERENCES [1] WHERE2 Partners. Deliverable D1.1b: Scenarios and parameters. Deliverable FP7ICT248894, WHERE2, 2011. [2] WHERE1 Partners. Deliverable D4.1: Measurements of location-dependent channel features. Deliverable FP7-ICT-2009-4, WHERE1, 2008. [3] WHERE2 Partners. Deliverable D2.3: Intermediate report on self-learning positioning using inferred context information. Deliverable FP7–ICT–248894, WHERE2, 2011. [4] WHERE2 Partners. Deliverable D1.8:final report on the where2 channel model. Deliverable FP7–ICT–248894, WHERE2, 2013. [5] WHERE2 Partners. Deliverable D1.5: Ray-tracing tools for dynamic positioning. Deliverable FP7-ICT248894, WHERE2, 2011. [6] http://www.pylayers.org.

34 / 146

ICT–248894 WHERE2

A

D2.6

A PPENDIX

The appendix contains a collection of articles and reports with detailed information to the summaries of the different sections in this deliverable, which are results of the WHERE2 project. Table 7 lists titles of the following sections in the appendix. Table 7: Overview of the collection of papers and reports. Appendix A.1 A.2 A.3 A.4 A.5 A.6 A.7 A.8 A.9 A.10 A.11 A.12

A.13

A.14 A.15 A.16

Title

Page

Context-Awareness and Self-Localization in Wireless Networks: The WHERE2 Proposals Technical Report on Simulations of the Synthetic Environment PyLayers: An Open Source Dynamic Simulator for Indoor Propagation and Localization Efficient Ray Tracing Tool for UWB Propagation and Localization Modeling Ray Tracing-Based Radio Propagation Modeling for Indoor Localization Purposes Model-based Evaluation of Location-based Relaying Policies in a Realistic Mobile Indoor Scenario Common Scenario Simulation Storage Rectangular Room Dimensions Estimation Using Narrowband Signal and Sectorized Antennas Self-Positioning and mapping of rectangular rooms with sectorized narrowband antennas Anchor-less Self-Positioning in Rectangular Room Based on Sectorized Narrowband Antennas Joint Anchor-less Tracking and Room Dimensions Estimation through IR-UWB Peer-to-peer Communications Benefits from Cooperation in Simultaneous Anchor-less Tracking and Room Mapping based on Impulse Radio Ultra Wideband Devices Cooperative Anchor-less Tracking and Room Mapping with Independent Computing Central Nodes and Application to Ultra Wideband Ray-Tracing Simulations Extraction of Context-Aware Features and Localization from Indoor-to-Outdoor Data Hidden Markov Model based Mobility Learning for Improving Indoor Tracking of Mobile Users Directional Hidden Markov Model for Indoor Tracking of Mobile Users and Realistic Case Study

35 / 146

36 45 58 64 70 76 82 88 95 101 108 114

120

125 134 140

ICT–248894 WHERE2

A.1

D2.6

Context-Awareness and Self-Localization in Wireless Networks: The WHERE2 Proposals

Igor Arambasic, Javier Casajus, Benoit Denis, Mathieu des Noes, Mohamed Laaraiedh, Bernard Uguen, Julien Stephan, Yves Lostanlen, Marios Raspopoulos, Stavros Stavrou, Jimmy J. Nielsen, Tatiana K. Madsen, Diogo M. Condeco, Ronald Raulefs, Context-Awareness and Self-Localization in Wireless Networks: The WHERE2 Proposals, ICT-Future Network and MobileSummit 2012, Berlin,Germany, July 2012

36 / 146

Future Network and MobileSummit 2011 Conference Proceedings Paul Cunningham and Miriam Cunningham (Eds) IIMC International Information Management Corporation, 2011 ISBN: 978-1-905824-23-6

Poster Paper

Context-Awareness and Self-Localization in Wireless Networks: The WHERE2 Proposals Igor ARAMBASIC1 , Javier CASAJUS1 , Benoˆıt DENIS2 , Mathieu DES NOES2 , Mohamed LAARAIEDH3 , Bernard UGUEN3 , Julien STEPHAN4 , Yves LOSTANLEN4 , Marios RASPOPOULOS5 , Stavros STAVROU5 , Jimmy J. NIELSEN6 , Tatiana K. MADSEN6 , Diogo M. CONDEC ¸ O7 , Ronald RAULEFS8 1 ETSI de Telecomunicaci´on, UPM, Av. Complutense 30, Madrid, 28040, Spain 2 CEA-Leti Minatec Campus, 17 Rue des Martyrs, Grenoble, 38054, France 3 Universit´e Rennes 1 - IETR UMR 6164, Av. du G´en´eral Leclerc, Rennes, 35042, France 4 SIRADEL, 3 All´ee Adolphe Bobierre, Rennes, 35043, France 5 Sigint Solutions Ltd, 231 Strovolos Av., Nicosia, 2048, Cyprus 6 Aalborg University, Fredrik Bajers vej 7, Aalborg, Denmark 7 Portugal Investigac¸a˜ o Explorat´oria, Rua Eng. Pinto Basto, Aveiro, 3810-106, Portugal 8 German Aerospace Center - DLR, Oberpfaffenhofen, Wessling, 82234, Germany Tel:+49 8153 282874, Email: [email protected] Abstract: This paper provides an overview of current research investigations carried out in the frame of the ICT WHERE2 project (FP7-ICT 248894, [1]), regarding mobility awareness, self-localization and environment characterization based on wireless communication networks. A first set of proposed techniques aims at retrieving the shape and the physical properties of indoor environments out of transmitted radio signals, while performing anchor-less localization. A second group of solutions concerns mobility self-learning, for the purpose of enhancing tracking functionalities. We also explain herein how the overall WHERE2 methodology enables a fair performance assessment of the designed algorithms, with the definition of a realistic evaluation framework. The latter comprises a selection of typical indoor environments, for which refined layout descriptions, real (Ultra) Wide Band ((U)WB) and Narrow Band (NB) measurements, together with site-specific ray-tracing simulations, are available. Keywords: Anchor-Less Localization, Context-Awareness, Indoor Mapping, Mobility, Ray-Tracing Simulations, Self-Learning, Tracking Filters

1. Introduction Context-awareness is clearly seen today as a major stake and enabling feature of future wireless networks. The so-called context is usually understood in a wide sense, which may concern physical elements of geometry, connectivity conditions or mobility aspects. Accordingly, the new concept of wireless context-awareness encompasses various functionalities such as environment characterization or mapping (i.e. retrieving a room volume, a building layout or walls materials), anchor-less self-localization and tracking (i.e. locating mobile terminals in their nearby environment and/or with respect to other mobile fellows with no fixed references) or mobility self-learning (i.e. inferring and predicting movement patterns and activity habits out of reconstructed trajectories). Ideally, these functionalities will be ensured out of standard radio communications and will enable a variety of applications such as context-based commercial services, c WHERE2 Copyright

www.FutureNetworkSummit.eu/2011

1 of 8

network optimization, assisted personal navigation in unknown environments. In the general context of heterogeneous and cooperative networks, the WHERE2 project (FP7ICT 248894, [1]) has been partly investigating key aspects of such functionalities, with specific research activities devoted to self-learning positioning using inferred context information. In this paper, we present a few illustrating examples regarding the joint estimation of room dimensions/volume and mobile locations, based on cooperative Ultra Wideband (UWB) links or mono-static channel sounding using sectorized antennas, as well as new techniques for learning users mobility patterns and enhancing tracking, e.g. by training Hidden Markov Models (HMM). The rest of the paper is organized as follows. The details of the synthetic test environment used for simulations are given in section 2. Then in section 3., we present new algorithms enabling simultaneous room mapping and self-localization, as well as preliminary related results based on both synthetic data and deterministic Ray-Tracing simulations. Section 4. subsequently introduces concepts related to mobility assessment, illustrating practically how to infer and to use movement patterns to improve tracking performances, while the last section is reserved for conclusions and discloses a few working perspectives on the related topics.

2. Synthetic Environment and Common Evaluation Framework The WHERE2 synthetic test environment is based on measurements collected during the WHERE1 project [2], which are complemented by deterministic ray-tracing simulations of Channel Impulse Responses (CIR) and a detailed 3D description of the site geometry. Simulations are calculated by using three distinct ray-based prediction tools: PyRay for Impulse Radio - Ultra Wideband (IR-UWB) signals, Volcano for Wide Band (WB) signals and 3DTruEM for Narrow Band (NB) signals [3]. The considered scenario is a gridbased representation of an indoor building environment where the user moves between rooms through doorways and corridors. Additionally this will be completed in turn by a 3D outdoor representation, intended for the extraction of context-aware features from indoor-to-outdoor channel data. The resulting data consist of CIRs simulated on a grid of 51 pseudo-Access Points (AP) x 363 mobile stations disseminated in the common indoor environment (i.e. the SIRADEL building), as illustrated in Fig. 1-a). On the selected simulation grid (chosen for providing sufficiently representative geometrical and channel configurations), the red points represent the transmitters while the blue points represent the receivers. The other colored elements highlight differences between constituting materials (for both the building structure and furniture pieces). For all the Tx-Rx links defined in Fig. 1-a) the synthetic CIRs are generated for UWB, WB and NB bandwidths. Fig. 2 and 3 show examples of UWB and NB CIRs obtained in the synthetic environment. Out of these CIRs, different location/context-radio parameters can be extracted, such as Received Signal Strength Indicators (RSSIs), multipath Times of Arrival (TOAs), (Average) Power delay Profiles ((A)PDPs), channel delay spread, or mean excess delay, etc. Accordingly, these CIRs are used as common inputs to validate the mapping and localization algorithms while the precise layout description is helpful to evaluate mobility learning and/or map-aided localization algorithms. Additionally, since the extraction of context-aware features is not constrained to indoor-to-indoor scenarios, the synthetic environment comprises indoor-to-outdoor information as well. An example is given in

c WHERE2 Copyright

www.FutureNetworkSummit.eu/2011

2 of 8

Figure 1: Detailed 3D description of the synthetic indoor environment with its measurement grid (Part of the SIRADEL building 1st floor) (a) and 3D overview of the indoor-to-outdoor environment (Overall SIRADEL building in yellow) (b).

Figure 3: NB CIRs with both omnidirecFigure 2: UWB CIRs with omnidirectional anten- tional and directional antennas in various nas for one Tx position and different Rx positions. directions for fixed Tx-Rx positions. Fig. 1-b), incorporating the detailed SIRADEL building in its nearby outdoor environment. Hence, by setting an AP inside the building and a terminal moving along an outdoor trajectory in the close building vicinity, simulated indoor-to-outdoor CIRs will be used to infer building features from the outside (e.g. coarse floor layout, presence of internal/external openings) offering complementary approaches to the indoor mapping solutions introduced hereafter.

3. Environment Mapping and Anchor-less Tracking One goal here is to estimate the 2D dimensions (and possibly the volume) of a rectangular room and to simultaneously localize devices communicating in this room in the absence of infrastructure. For this sake, one idea is to interpret the structure of the received radio signals. As an example, the multipath echoes issued from simple sequences of electromagnetic interactions with the environment (e.g. single-bounce reflections) are usually assumed to be energetic, spatially correlated and to convey constructive information with respect to geometry (including mobile locations). A first solution put forward in the frame of WHERE2 thus relies on the highresolution IR-UWB technology to estimate the TOAs of such strong multipath components between various pairs of cooperative devices (i.e. operating within groups of more than 2 nodes), considering jointly the indoor mapping and mobile tracking problems. As a first trivial step, for each pair-wise link, the excess delays {τj }j=0..4 associated with the 4 single-bounce reflected paths are theoretically related to the wall dimensions (Dx , Dy ) and to the coordinates of the communicating nodes in this room {(xm , ym ), (xn , yn )} (e.g. See Fig. 4-a)). However, in real-life systems, the measured delays {ti }i=0..4 isc WHERE2 Copyright

www.FutureNetworkSummit.eu/2011

3 of 8

sued after TOA estimation still need to be mapped onto the a priori modeled {τj }j=0..4 (i.e. according to the arbitrary wall numbering). Assuming systematic Line of Sight (LOS) channel configurations in the room, the first measured delay t0 can be systematically associated with the direct path and hence mapped onto τ0 . But there are still 24 remaining mapping combinations to be tested for each link, what becomes even worse when considering all the cooperative links simultaneously. This data association problem can be solved during the initialization phase (i.e. considering that the initial mapping guess can remain unchanged afterwards) and/or continuously updated in the steady-state tracking regime. In compliance with the cooperative framework, a specific state initialization procedure based on the minimization of quadratic cost functions has been proposed. The underlying idea is that the most probable labeling option would generate somehow the lowest residual after minimization among all the possibilities. As regards to continuously updated data association (i.e. in the steady-state regime), one can select the set of TOA measurements that is the nearest to the set predicted by the tracking filter. Besides, a new cooperative tracking filter structure coping with at least three cooperative nodes has been introduced, the state vector containing the coordinates of all the involved devices, along with the room dimensions.

Figure 4: Single-bounce reflection model in a rectangular room with arbitrary wall numbering (2 nodes only) (a) and CDF of locations and room dimensions estimation errors in a 5mx10m room with 3 cooperative UWB nodes and a 4ns multipath TOA estimation uncertainty (b). The performance of the proposed algorithm was assessed through simulations first, considering synthetic multipath TOA measurements first, averaging the estimation errors over multiple mobile trajectories in a 5mx10m room. For comparison with the proposed optimization-based scheme, the results obtained with a perfect state initialization were also recorded. The CDF of estimation errors affecting both the location and the room dimensions are shown on Fig. 4-b), for moderate mobility (i.e. maximum speed < 0.5m.s−1 ). In spite of a purely blind initialization and reasonably conservative TOA precision levels (i.e. compatible with the time resolution usually available at real UWB receivers, e.g. 4ns in the shown example), the anchor-less location performance appears to be on the order of that achieved by conventional anchors-based solutions, say with a median error better than 1m (i.e. on the order of 0.5m), what looks compliant with most of indoor application requirements. Contrary to the previous UWB method, hardware constraints are mostly imposed on antenna design in a second approach. The bandwidth is thus reduced down to 100MHz (NB domain) and the receiver isotropic antenna is replaced by eight sectorized antennas with 45 ◦ aperture in the horizontal plane. The total coverage of these antennas correc WHERE2 Copyright

www.FutureNetworkSummit.eu/2011

4 of 8

sponds to the isotropic one. The system consists of two radio transceivers, both capable of full duplex operations. This facilitates CIR measurements at the same place where the signal is transmitted (generated), commonly known as self-to-self CIRs. When representing a rectangular room as the reference geometrical model, walls are seen as reflecting lines in a two-dimensional space, while corners are characterized by double reflections. Hence, by combining the information obtained by self-to-self CIR at each antenna, the presence of walls and corners can be evaluated utilizing round trip time (RTT) measurements. This hypothesis is true if we consider four sectors of Rx antenna to be orthogonal to the walls, meaning their TOA calculations correspond to radar readings. On the other hand, four remaining Rx antenna sectors are pointing to the corners and, since their signal include double-reflections, the corresponding distance readings must be systematically corrected. The correction should be smaller for the corners that are far away from the transceiver and larger if the transceiver is located closer to the corner. This correction parameter, labeled as C, is set empirically between 0.7 and 1 (e.g. with C=0.7 fitting to raw distance measurements between 0m and 1.5m or C=0.8 between 3 and 4.5m). In order to obtain the room volume, a deeper a priory analysis of the CIRs is required. Accordingly a common absorption coefficient is determined for the entire test environment as 0.492, while path loss model parameters are set to L0 =-35.85dB (i.e. reference path loss at 1m) and n=1.7 (path loss exponent). The procedures for calculating the absorption coefficient and the path loss parameters characterizing the SIRADEL building environment are detailed in [4]. The algorithm starts by calculating the distance between transceivers, out of isotropic standard CIRs, using the a priori calibrated path loss model. Afterwards, the Rx sectorized CIRs of both transceivers are used for detecting the direction of the emitter, which lies inside the antenna sector with the highest power level. The same calculation is done on the other transceiver and the resulting angle is the average of the two readings. The resulting precision of angle classification is 22.5 ◦ . After these two steps the

Figure 5: Example of calculation for a left wall out of 12 possible points transceivers are positioned with respect to each-another, enabling the data obtained out of self-to-self CIRs to be constructive. The third step consists of locating the RTT of the strongest peak for the self-to-self CIRs of the sectorized Rx antennas, at both positions. In the NB case (i.e. 100MHz here), TOA values are multiples of 10ns, which matches any round-trip distance to the wall to be a multiple of 1.5m. Hence, if the c WHERE2 Copyright

www.FutureNetworkSummit.eu/2011

5 of 8

RTT of the orthogonal section is 10ns, the wall is positioned between 1.5m and 3m from the transceiver. Eventually these readings (i.e. 1.5m and 3m) lead to two points that the corresponding Rx sector-antenna would produce. Hence, when determining the wall position, we part with 12 observations for each wall, as seen in Fig. 5. Four correspond directly to orthogonal wall observations, and eight correspond to the two corners. These 12 observations are then projected on the straight line and the median of all projections produces the wall position estimate.

Figure 6: CDFs of relative errors on estimated room dimensions X and Y (a) and estimated room volume (b) The CDFs of the relative errors affecting the estimated Dx and Dy dimensions are depicted in Fig. 6-a), for one 5.1mx7.3m room in the environment described in Section 2. (i.e. a reception hall consisting of three doors, with a metallic bookcase, one large wooden desk and one table). The room is not perfectly rectangular and, as such, could be considered as rather hostile. Nevertheless, the obtained CDF shows that in more than 80% of the calculations the relative error of the obtained dimensions is bellow 15%, while in 50% of calculations the error would be lower than 11%. After obtaining the two room dimensions, these can be used together with the common absorption coefficient for determining the room height and consequently the room volume. The obtained performance, shown in Fig. 6-b), is in line with the relative errors obtained for Dx and Dy , meaning that no significant error was added in this step.

4. Mobility Patterns Learning In the tracking of mobile users, any prior knowledge about the movement properties or behavior may be helpful. Typically, data are smoothened or filtered if the trajectory is continuous and is unlikely to make sharp turns. Hence prior information about walls and obstacles can for instance contribute to improve tracking accuracy through forecasting of user turns. An alternative approach put forward in WHERE2 is based on exploiting historic mobility traces of the users to enhance next positions prediction. In order to obtain a realistic mobility representation a self-learning model is realized as a HMM. The training data (gathered on-line or off-line) are used to adjust transition probabilities, as seen in Fig. 7-a). Afterwards this mobility model can be incorporated into a filtering solution (e.g. See Fig. 7-b)), e.g. based on the Viterbi algorithm that we refer to as an adaptive Viterbi algorithm. In the off-line learning approach, the training phase and construction of the matrix with transition probabilities is realized a priori, before feeding the tracking filter. In the on-line version, any filter output is used to update the transition matrix concurrently. c WHERE2 Copyright

www.FutureNetworkSummit.eu/2011

6 of 8

Figure 7: Projection of a movement area onto a Markov chain (a) and general framework of mobility learning for enhanced tracking (b). Preliminary simulations were carried out to evaluate the performance of the proposed solutions. Fig. 8 compares the mean location errors obtained after performing a local smoothing of the estimated movement (1), HMM-based filtering without mobility model knowledge (i.e. assuming all directions are equiprobable) (2); HMM-based filtering assuming a straight movement of the object (3); and HMM-based filtering using a self-learned mobility model (4). These results indicate that the adaptive approach based on a self-learned mobility can improve significantly the estimation accuracy. 1.4 Moving average (1) Viterbi (2) Directional Viterbi (3) Adaptive Viterbi (4)

1.2

Error [grid points]

1

0.8

0.6

0.4

0.2

0

offline

online

Figure 8: Performance comparison of different tracking approaches assisted by offline and online mobility learning. According to a second WHERE2 proposal, we consider enhancing a fingerprinting localization system based on WiFi RSSI readings (collected with respect to a few APs), by incorporating historical information regarding past users movements in the building. More particularly, in order to select one single reliable user position (out of the prior database), which is clearly hazardous in classical RSSI-based fingerprinting approaches due to large random power fluctuations (e.g. small-scale fading under mobility, shadowing, device orientation, etc.), the movement history can be analyzed and statistically described. Indeed the movements inside a building tend to be repeated, so that identifying mobility patterns is helpful not only to limit the space of results (i.e. possible locations in the fingerprinting database), but also to infer plausible destinations or directions to assist location predictions. As a first step of this investigation, real user data, including the position information, were recorded for different long-term time windows in the frame of a WiFi measurements campaign. The users movement patterns were subsequently identified. The movement c WHERE2 Copyright

www.FutureNetworkSummit.eu/2011

7 of 8

identification process and storage into the designed database is represented in Fig. 9. Following a similar approach as in Fig. 7-a), the probabilities of visiting a cell for a given user can be calculated based on the history of recent locations, which can be either the latest location estimates or an a priori known location (e.g. passing through an RFID gate). As another example, considering that the user is inside a corridor and that he tends follow the shortest way between two points while walking, one can assume that the movement is constrained to a linear route (forward and backwards). Besides these preliminary efforts regarding movement analysis, using the real data mentioned above, new algorithms relying on these concepts are currently developed and tested within fingerprinting-based positioning.

Figure 9: Database structure considered for the analysis of users mobility patterns within a real WiFi indoor measurement campaign.

5. Conclusion and Perspectives In this paper we have accounted for current efforts realized in the frame of the WHERE2 project to provide new context-aware functionalities into wireless networks. A special focus has been put on indoor mapping, anchors-less localization and mobility learning. Although the concepts and solutions discussed above are still non-definitive, the preliminary results obtained so far are encouraging and look already reasonably competitive with respect to application needs. In terms of perspectives, several axes of investigation have been identified for the near future. As regards to simultaneous localization and mapping, more realistic receiver models and channel profiles (e.g. incorporating the UWB Ray-Tracing simulations) shall be considered, and new techniques will be proposed to infer building features from indoor-to-outdoor radio transmissions. As regards to mobility learning, the feasibility of the proposed schemes will be assessed based on real data in the common evaluation framework, coupling the new functionalities with WHERE2 algorithms such as map-aided localization or mobility-based handover.

References [1] http://www.ict-where2.eu/. [2] “Measurements of location-dependent channel features,” tech. rep., Deliverable D4.1 of the WHERE1 Project (ICT-217033), Oct. 2008. [3] “Ray-tracing tools for dynamic positioning,” tech. rep., Deliverable D1.5 of the WHERE2 Project (ICT-248894), July 2011. [4] “Self-learning positioning using inferred context information (intermediate report),” tech. rep., Deliverable D2.3 of the WHERE2 Project (ICT-248894), Dec. 2011.

c WHERE2 Copyright

www.FutureNetworkSummit.eu/2011

8 of 8

ICT–248894 WHERE2

A.2

D2.6

Technical Report on Simulations of the Synthetic Environment

Mohamed Laaraiedh, Bernard Uguen, Marios Raspopoulos, and Julien Stephan

45 / 146

1

Technical Report On Simulations of the Synthetic Environment Mohamed Laaraiedh, Bernard Uguen, Marios Raspopoulos, and Julien Stephan

I. I NTRODUCTION During the first year of the WHERE2 project, partners within the task 2.3 expressed the need for Channel Impulse Responses (CIR) for several radio downlinks in an indoor environment where the transmitters and receivers are not necessarily located inside the same main walls. Based on these requirements presented in Table I, the Ultra WideBand (UWB) channel measurements collected during the WHERE1 project [1] were pointed out as a possible starting point to build a test environment. This measurement campaign was performed using 4 receivers and a transmitter moving on a grid of 302 locations spread along the SIRADEL premises. However, these measurements characterize only one specific bandwidth (i.e. UWB) and do not cover enough scenarios (in term of radio links) in order to fulfill all the partners’ needs. Ray-based techniques appear thereby a valuable solution to complement these measurements by deterministic simulations of impulse responses. Simulations were thus calculated by the three partners of the WHERE2 project who have been developing and enhancing their own ray-based solution [2]: • PyRay ray tracer to obtain Utra Wide Band (i.e. > 500MHz) simulations • Volcano technology based ray tracer Wide Band (i.e. ∼ 500MHz) simulations • 3DTruEM ray tracer to obtain Narrow Band (i.e. < 100MHz) simulations In addition, a complementary method based on the extraction of context-aware features from indoor-to-outdoor channel data is also studied. Its elaboration and evaluation necessarily rely on realistic environment and channel realizations. In that purpose, the 3D building model (i.e. the SIRADEL premises) is extended to include a fine representation of the surroundings, to be later used by the indoor-to-outdoor ray-tracing model developed in WP1 [3] in order to get realistic indoor/outdoor channel realizations along new indoor-to-outdoor radio links. Analysis of these indoor/outdoor channel realizations will then process to identify if building features (walls and apertures in particular) can be inferred from the outside, which could finally help the localization of the indoor terminals. T2.3 Partner UPM

Main Purpose Map Construction

SIG UR1

Map Construction Map Construction

CEA

Map Construction and Positioning enhancement Tracking enhancement Indoor/Outdoor case study

AAU SIR

Bandwidth 100MHz, 500MHz < 100MHz UWB (> 500MHz) UWB (> 500MHz) Any 100MHz, 500MHz

Required parameters CIR CIR, RSSI, AoA CIR, RSSI, TOA CIR RSSI and TOA CIR

TABLE I: T2.3 Partners’ requirements: All the partners require peer-to-peer measurements described by a grid of at least 1x1m.

II. C REATION OF AN INDOOR SYNTHETIC TEST ENVIRONMENT The indoor synthetic environment defined within T2.3 is plotted in Figure 1 and 2 for respetively 2D and 3D landmarks. The points represent the grid chosen for performing simulations: the red points represent the transmitters while the blue points represent the receivers. These coordinates of these transmitters and receivers are respectively shown in Table IV and Table III. The different components of the environments (i.g. doors, walls, etc.) are represented in different colours to highlight the difference between materials. The properties of these materials, shown in Table II, are taken into consideration by the different ray tracers in order to perform realistic simulations.

2

Fig. 1: 2D description of the synthetic environment.

Fig. 2: 3D description of the synthetic environment.

Material Category WALLS WOODEN DOOR INSULATION PLASTERBOARD 14CM PLASTERBOARD 7CM CONCRETE 15CM3D CONCRETE 20CM3D CONCRETE 7CM3D CONCRETE 25CM3D 3D WINDOW GLASSES WOODEN TABLES DESKS FLOOR TOP CEILING CEILING Metallic

Plastic

Description Internal walls assumed to be made of plaster Wooden Doors Perimeter pillars covered with black insulation 14cm plasterboard 7cm plasterboard 15cm–thick concrete walls in the entrance of the building 20cm–thick concrete walls along the perimeter of the building 7cm–thick concrete walls. One on the very left and one on the very right of the building 25cm–thick wall Glasses along the perimeter of the building and one glass inside the building Wooden tables and desks 50cm–thick floor 50cm–thick ceiling Intermediate plaster ceiling (40 cm below the top ceiling) All the metallic objects of the building including copy machines, refrigerator and printers Plastic doors of the metallic cupboards

0.1

Complex Permittivity Real Part 2.5

Complex Permittivity Imag. Part –0.3

0.1 0.2

2.5 4

–0.03 –0.4

0.14 0.07 0.15

2.5 2.5 7

–0.3 –0.3 –0.8

0.2

7

–0.8

0.07

7

–0.8

0.25 0.05

7 6

–0.8 –0.05

0.01 0.5 0.5 0.005

2.5 5 5 2.5

–0.03 –0.4 –0.4 –0.3

0.005

1

–1000000

0.1

4

–0.4

Thickness (m)

TABLE II: Complex Permittivity parameters calibrated for 3DTruEM using the M2 Measurement campaign of WHERE-1

3

Rx ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Grid Position (0.5,1) (0.5,2) (0.5,3) (0.5,4) (0.5,5) (0.5,6) (0.5,7) (0.5,8) (0.5,9) (0.5,10) (0.5,11) (1.5,1) (1.5,2) (1.5,3) (1.5,4) (1.5,5) (1.5,6) (1.5,7) (1.5,8) (1.5,9) (1.5,10) (1.5,11) (2.5,1) (2.5,2) (2.5,3) (2.5,4) (2.5,5) (2.5,6) (2.5,7) (2.5,8) (2.5,9) (2.5,10) (2.5,11) (3.5,1) (3.5,2) (3.5,3) (3.5,4) (3.5,5) (3.5,6) (3.5,7) (3.5,8) (3.5,9) (3.5,10) (3.5,11) (4.5,1) (4.5,2) (4.5,3) (4.5,4) (4.5,5) (4.5,6) (4.5,7) (4.5,8) (4.5,9) (4.5,10) (4.5,11) (5.5,1) (5.5,2) (5.5,3) (5.5,4) (5.5,5)

Rx ID 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122

Grid Position (5.5,8) (5.5,9) (5.5,10) (5.5,11) (6.5,1) (6.5,2) (6.5,3) (6.5,4) (6.5,5) (6.5,6) (6.5,7) (6.5,8) (6.5,9) (6.5,10) (6.5,11) (7.5,1) (7.5,2) (7.5,3) (7.5,4) (7.5,5) (7.5,6) (7.5,7) (7.5,8) (7.5,9) (7.5,10) (7.5,11) (8.5,1) (8.5,2) (8.5,3) (8.5,4) (8.5,5) (8.5,6) (8.5,7) (8.5,8) (8.5,9) (8.5,10) (8.5,11) (9.5,1) (9.5,2) (9.5,3) (9.5,4) (9.5,5) (9.5,6) (9.5,7) (9.5,8) (9.5,9) (9.5,10) (9.5,11) (10.5,1) (10.5,2) (10.5,3) (10.5,4) (10.5,5) (10.5,6) (10.5,7) (10.5,8) (10.5,9) (10.5,10) (10.5,11) (11.5,1)

Rx ID 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184

Grid Position (11.5,4) (11.5,5) (11.5,6) (11.5,7) (11.5,8) (11.5,9) (11.5,10) (11.5,1) (12.5,1) (12.5,2) (12.5,3) (12.5,4) (12.5,5) (12.5,6) (12.5,7) (12.5,8) (12.5,9) (12.5,10) (12.5,11) (13.5,1) (13.5,2) (13.5,3) (13.5,4) (13.5,5) (13.5,6) (13.5,7) (13.5,8) (13.5,9) (13.5,10) (13.5,11) (14.5,1) (14.5,2) (14.5,3) (14.5,4) (14.5,5) (14.5,6) (14.5,7) (14.5,8) (14.5,9) (14.5,10) (14.5,11) (15.5,1) (15.5,2) (15.5,3) (15.5,4) (15.5,5) (15.5,6) (15.5,7) (15.5,8) (15.5,9) (15.5,10) (15.5,11) (16.5,1) (16.5,2) (16.5,3) (16.5,4) (16.5,5) (16.5,6) (16.5,7) (16.5,8)

Rx ID 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246

Grid Position (16.5,11) (17.5,1) (17.5,2) (17.5,3) (17.5,4) (17.5,5) (17.5,6) (17.5,7) (17.5,8) (17.5,9) (17.5,10) (17.5,11) (18.5,1) (18.5,2) (18.5,3) (18.5,4) (18.5,5) (18.5,6) (18.5,7) (18.5,8) (18.5,9) (18.5,10) (18.5,11) (19.5,1) (19.5,2) (19.5,3) (19.5,4) (19.5,5) (19.5,6) (19.5,7) (19.5,8) (19.5,9) (19.5,10) (19.5,11) (20.5,1) (20.5,2) (20.5,3) (20.5,4) (20.5,5) (20.5,6) (20.5,7) (20.5,8) (20.5,9) (20.5,10) (20.5,11) (21.5,1) (21.5,2) (21.5,3) (21.5,4) (21.5,5) (21.5,6) (21.5,7) (21.5,8) (21.5,9) (21.5,10) (21.5,11) (22.5,1) (22.5,2) (22.5,3) (22.5,4)

Rx ID 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308

Grid Position (22.5,7) (22.5,8) (22.5,9) (22.5,10) (22.5,11) (23.5,1) (23.5,2) (23.5,3) (23.5,4) (23.5,5) (23.5,6) (23.5,7) (23.5,8) (23.5,9) (23.5,10) (23.5,11) (24.5,1) (24.5,2) (24.5,3) (24.5,4) (24.5,5) (24.5,6) (24.5,7) (24.5,8) (24.5,9) (24.5,10) (24.5,11) (25.5,1) (25.5,2) (25.5,3) (25.5,4) (25.5,5) (25.5,6) (25.5,7) (25.5,8) (25.5,9) (25.5,10) (25.5,11) (26.5,1) (26.5,2) (26.5,3) (26.5,4) (26.5,5) (26.5,6) (26.5,7) (26.5,8) (26.5,9) (26.5,10) (26.5,11) (27.5,1) (27.5,2) (27.5,3) (27.5,4) (27.5,5) (27.5,6) (27.5,7) (27.5,8) (27.5,9) (27.5,10) (27.5,11)

Rx ID 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363

Grid Position (28.5,3) (28.5,4) (28.5,5) (28.5,6) (28.5,7) (28.5,8) (28.5,9) (28.5,10) (28.5,11) (29.5,1) (29.5,2) (29.5,3) (29.5,4) (29.5,5) (29.5,6) (29.5,7) (29.5,8) (29.5,9) (29.5,10) (29.5,11) (30.5,1) (30.5,2) (30.5,3) (30.5,4) (30.5,5) (30.5,6) (30.5,7) (30.5,8) (30.5,9) (30.5,10) (30.5,11) (31.5,1) (31.5,2) (31.5,3) (31.5,4) (31.5,5) (31.5,6) (31.5,7) (31.5,8) (31.5,9) (31.5,10) (31.5,11) (32.5,1) (32.5,2) (32.5,3) (32.5,4) (32.5,5) (32.5,6) (32.5,7) (32.5,8) (32.5,9) (32.5,10) (32.5,11)

TABLE III: Receiver Coordinates. The origin [0,0] is located at the lowest-left corner of the building

4

Rx ID 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51

Grid Position (0.5,11) (2.5,4) (2.5,9) (4.5,1) (4.5,5) (4.5,7) (4.5,11) (7.5,3) (7.5,5) (8.5,4) (9.5,11) (11.5,4) (12.5,5) (13.5,1) (13.5,8) (14.5,4) (15.5,10) (16.5,5) (17.5,7) (18.5,1) (18.5,11) (19.5,1) (19.5,4) (20.5,5) (21.5,9) (22.5,10) (23.5,3) (23.5,8) (24.5,2) (24.5,4) (24.5,11) (25.5,3) (25.5,7) (28.5,1) (28.5,7) (28.5,8) (28.5,11) (29.5,4) (29.5,6) (29.5,10) (29.5,11) (30.5,4) (30.5,10) (31.5,2) (31.5,4) (31.5,9) (31.5,10) (32.5,1) (32.5,7) (32.5,8) (32.5,11)

TABLE IV: Transmitters Coordinates. The origin [0,0] is located at the lowest-left corner of the building

5

III. UWB SIMULATIONS OF THE INDOOR SYNTHETIC ENVIRONMENT In this section the UWB based simulations of the synthetic environment, are presented and detailed. These simulations are performed using the ray-tracing tool “PyRay” developed by the University of Rennes 1 (UR1) which is a tool developed by the University of Rennes 1. PyRay is a Python written simulator which allows indoor radio wave propagation modelling for multi-standard systems, including UWB. It is based on the combination of a 3D ray-tracing method with geometrical optics and uniform theory of diffraction. The complete description of PyRay is given in [2]. In order to perform simulation with the PyRay tool, a detailed 3D site geometry, material properties, transmitter and receiver locations, antenna properties and orientation are required. In order to perform simulation with ray tracer tools, a detailed 3D site geometry, material properties, transmitter and receiver locations, antenna properties and orientation have been used to parametrize the simulations. Moreover, a precise description of the shape of the pieces of furniture is also available to make possible the generation and posterior emulation of plausible mobility patterns (and hence mobile radio traces) in the same environment. The obtained simulations are a set of Matlab files which store the different channel impulse responses (CIR) of all the defined links between Transmitters and Receivers in Figure 1. In addition to the CIR, the coordinates of Tx and Rx are also available. Figure 3 plots an example of obtained CIRs and Figure 4 plots the corresponding results in frequency domain. From these CIR, different radio parameters can be extracted such as RSSI, TOA, τRM S , τmean , etc.

10

30

50

Voltage (V)

20

0.0006 0.0004 0.0002 0.0000 0.0002 0.0004 0.000620

Rx 10 Rx 30 Rx 50

0

20

40

60 Time (ns)

80

100

120

140

Fig. 3: Examples of IR obtained in the synthetic environment.

10

30

50

Voltage (V)

20

50 100 150 200 250 300 350 400 45020

Rx 10 Rx 30 Rx 50

40

60

80 Time (ns)

100

120

140

Fig. 4: Examples of frequency domain CIR obtained in the synthetic environment. In many SLAM algorithms, it is necessary to identify the rays reflected on or transmitted across the wall which is to be reconstructed. The UWB simulations provide, in addition to the radio-parameters mentioned before, the possibility to identify one, a set, or all the rays which are implicated in the CIR. Figure 5 plot all the obtained rays between two points. As an example, the 5th strongest ray is plotted in Figure 6. Furthermore, for each ray different informations are available and can be exploited within T2.3 and WHERE2 project.

6

Fig. 5: Examples of obtained rays in the synthetic environment.

Fig. 6: Each ray can be identified and exploited individually. a) Format of the simulations files: The simulations results are presented into Matlab files (.mat Format) and are put on the BSCW server (Figure 7) for further exploitation by the partners. The results are saved as follows: • For each transmitter (idtx from 001 to 051), a zipped file (Tx[idtx].zip) is created which contains the simulations results of all radio links between the current transmitter and all the receiver (idrx from 001 to 343). • Inside each zipped file, 343 matlab files contains the simulations of all the radio links, the format of these matlab files is: cir-tx[idtx]-rx[idrx].mat • Each matlab file contains the following informations: – The coordinates of the Tx – The coordinates of the Rx – The CIR between the Tx and the Rx without taking into consideration the antenna – The time base of the CIR without taking into consideration the antenna – The CIR when taking into consideration the antenna (the antenna used in M1 measurements, see D4.1 from WHERE1 [1]) – The time base of the CIR when taking into consideration the antenna An example of processing of a matlab file is given in Figure 8. b) Parameters extracted from simulations: Different parameters can be extracted from the obtained CIR. Among this parameters, the following are the most important: • Distance between Tx and Rx q 2 2 2 d = (xT x − xRx ) + (yT x − yRx ) + (zT x − zRx ) •

where (xT x , yT x , zT x ) and (xRx , yRx , zRx ) are respectively the coordinates of the Tx and the Rx Time of flight τ0 = d/c

7

Fig. 7: Simulation files.

Fig. 8: Example of processing of simulation files.

•

Transmitted power Et = 10 log10

•

Z

T

δ 2 (t)dt

0

where δ(t) is the transmitted pulse and t is the time base RSSI: total received energy (Etot ) and The energy integrated over the strongest received path (Emax ) Z T Etot = 10 log10 r2 (t)dt 0

Emax = 10 log10

•

where r(t) is the CIR The Mean excess delay τmoy

Z

τM +(1−α)Ts

τM −αTs

R∞ 2 tr (t)dt = R0∞ 2 r (t)dt 0

r2 (t)dt

8

•

The Delay spread τrms =

•

sR ∞ 0

r2 (t)(t − τmoy )2 dt R∞ r2 (t)dt 0

Time of arrival of the first path (TOA) obtained by different techniques τˆth , τˆcum , τˆmax , τˆwin

•

Angle of arrival αtr = arctan 2(

yT x − yRx ) xT x − xRx

IV. NARROW-BAND SIMULATIONS OF THE SYNTHETIC INDOOR ENVIRONMENT Narrow-band simulations have also been carried out using Sigint Solutions Ray Tracing Simulator 3DTruEM. 3DTruEM is a full–3D fully deterministic Ray Tracing simulator which was developed in .NET programming languages by Sigint Solutions and is expected as an internal project to be launched as a commercial product of the company in the beginning of 2012. It is based on the Uniform Theory of Diffraction (UTD) and it is fully polarimetric (considers all the polarisations of the transmitted signal). Figure 9 shows a screen shot of this Simulator. More details regarding this simulator can be found in WHERE-2 Deliverable D1.6 [3].

Fig. 9: 3DTruEM Screenshot As in the case of the UWB simulations in section III the full 3D model of the synthetic environment has been generated using 3DTruEM built–in CAD designer by using the 2D DXF file of the SIRADEL environment and also the clutter (furniture) described in Deliverable D1.1 [4] of the WHERE–1 Project. A screen shot of the developed model is shown in figure 10. A detailed morphological description of the synthetic environment exists since the electrical properties of the environment facets have been calibrated using the M2 Measurement campaign of the WHERE-1 project (see D11 WHERE–1 [4]). The calibration procedure can be found in WHERE–1 D4.2 [5]. The extracted calibrated electrical parameters are shown in Table II. The simulations have been carried out at 3.5GHz (this is the frequency that the Ray Tracing Simulator was calibrated for) and a bandwidth of 100M Hz considering an unlimited number of receivers, an unlimited number of refractions and up to 1 diffraction. The receiver sensitivity was set as a terminating condition for every propagating ray launched from the transmitters. 51 isotropic Transmitters have been defined and their transmitting power was set to 10dBm. Also 363 isotropic Receivers with Receiver sensitivity -110dBm have been defined (their coordinates are tabulated in Table IV). Simulations were carried out and the impulse responses for each receiver location and respect to each Transmitter was computed. All the results have been exported in MATLAB and a separate MATLAB structure was created for each Transmitter. Each datafile is save as a separate file having the following filename Narrow RT Sigint TXN where N indicates the number of the respective transmitter. The structure of the results is shown in Figure 11. It contains various details about the simulation parameters and the IR received at the receiver location an omnidirectional antenna but also the 8 IRs received by directional antennas rotated by 45 degrees each time. Figure 12 shows some typical IRs obtained for a specific transmitter receiver combination (TX1 and RX23) using omnidirectional antennas and directional antennas in various directions.

9

Fig. 10: 3D Model of the Synthetic Environment for Simulations using 3DTruEM

Fig. 11: MATLAB structure for every Transmitter

Fig. 12: Typical narrow–band Impulse Responses for a specific Transmitter-Receiver Combination for omnidirectional antenna and directional antenna in various directions

10

V. W IDE -BAND SIMULATIONS OF THE SYNTHETIC INDOOR ENVIRONMENT The resulting data consist of impulse responses simulated along a grid of 51 Access Points x 363 Mobile stations spread on a common indoor synthetic environment (i.e. the SIRADEL building) as illustrated in the Figure 13. In this section, the bandwidth is around 500MHz. The Table V summarizes the whole simulation parameters used by SIRADEL for the wideband simulations. Note that the ray-tracing used by SIRADEL for these simulations was beforehand calibrated from the indoor measurement campaign M2 collected during the WHERE1 project (i.e. channel sounder measurements collected at 3.5 GHz in the SIRADEL building): an adjustment of the dielectric properties of the internal partitions was notably done and a correction factor on the free-space loss was applied.

Fig. 13: Simulation set-up - 3D overview. Parameters Environment representation Transmitted power Maximum delay

Antenna type

Receiver sensitivity Height of Transmitters Height of Receivers

Description 2.5D Digital Building Model (DBM) of the SIRADEL building with furniture 10 dBm 300ns Isotropic antennas for Tx/Rx. Isotropic antenna for Tx and sector antenna for Rx with theoretical radiation pattern: 45 ◦ aperture in the horizontal plane. Omnidirectional radiation pattern in the vertical plane. Different azimuth: [0 ◦ , 45 ◦ , 90 ◦ , 135 ◦ , 180 ◦ , 225 ◦ , 270 ◦ , 315 ◦ ] -110dBm 1.5m 1.5m

TABLE V: SIRADEL Indoor Ray-tracing Solution – T2.3 Impulse Response simulation parameters

VI. C REATION OF A INDOOR - TO - OUTDOOR SYNTHETIC TEST ENVIRONMENT Investigations are conducted on the extraction of context-aware features from indoor-to-outdoor channel data. Construction of a new indoor-to-outdoor synthetic test environment is needed for this purpose. It consists of both geometrical and physical

11

data derived from a natural scene. SIRADEL uses a two steps process in order to create this indoor-to-outdoor synthetic test environment. First step consists in the creation a highly realistic map data of the environment, including representation of the ground, vegetation, urban furniture possibly, building shapes, building facade details (door and window apertures in particular), inbuilding partitions and in-building furniture. SIRADEL defines the following process to get the expected map data precision: 1) Getting or/and Creating 3D high-resolution outdoor map data of the environment (obtained mainly from aerial pictures) and architect floor plans of the buildings under test; 2) Collecting on field measurement data using a Mobile Mapping System. To this end, SIRADEL disposes of a specific van that integrates an advanced GPS, three 2D lasers and five cameras; 3) Post-processing the data collected in (2.) by using SIRADEL proprietary algorithms: a) Laser data are used to automatically rectify the building boundaries and align ground data (camera) on aerial data (3D high resolution map data). b) Image sequences are used to texture the buildings facades. c) Laser and image sequences are used conjointly to automatically detect the presence of the openings (windows and doors) and their precise locations; 4) Geo-locating and scaling precisely the architect plans obtained in (1.) according to the rectified boundary and the precise openings location of the buildings under test obtained in (3.); 5) Drawing manually the main internal partitions (i.e. strong and light walls) in order to get a realistic 3D indoor-to-outdoor map data representation. So far, three realistic indoor-outdoor map data have already been constructed. One of these environments is illustrated in Figure 14.

Fig. 14: Example of an indoor-outdoor environment already constructed – 3D overview. The second step aims at deriving physical data. For our investigations, physical data consist in simulated Channel Impulse Responses (CIRs) for indoor-to-outdoor radio links. One terminal is fixed inside a building and another terminal is moving along an outdoor trajectory in the building close vicinity. In order to produce such data, we are currently working on the elaboration of a novel site-specific channel model. This model is based on the SIRADEL ray-tracing tool presented in [3] and an advanced method to predict the indoor/outdoor radio wave propagation. VII. C ONCLUSION In this technical report, the simulations results of the indoor synthetic environment are presented. Three bandwidths are considered: UWB, Wide band and Narrow Band. These results are obtained using three different ray-tracers developed by three WHERE2 partners. The first investigations of these simulations results have shown that different parameters (radio and non-radio) can be extracted and should be very useful within WHERE2 localization algorithms. Furthermore, the simulations

12

allow the identification and the exploitation of each ray alone which is very interesting for SLAM algorithm, one goal of task T2.3. Similarly to the indoor synthetic environment, an indoor-to-outdoor synthetic environment is also being defined by SIRADEL in order to to extract context-aware features. R EFERENCES [1] [2] [3] [4] [5]

WHERE1 WHERE2 WHERE2 WHERE1 WHERE1

Partners. Partners. Partners. Partners. Partners.

Deliverable Deliverable Deliverable Deliverable Deliverable

d4.1: d1.5: d1.6: d1.1: d4.2:

Measurements of location-dependent channel features. Deliverable FP7-ICT-2009-4, WHERE1, 2008. Ray-tracing tools for dynamic positioning. Deliverable FP7-ICT248894, WHERE2, 2011. Ray–tracing tool for heterogeneous distributed positioning. Deliverable FP7-ICT248894, WHERE2, 2011. Definition of the where framework and scenarios. Deliverable FP7-ICT-2009-4, WHERE1, 2008. Fusion of ray tracing and channel measurements. Deliverable FP7-ICT-2009-4, WHERE1, 2008.

ICT–248894 WHERE2

A.3

D2.6

PyLayers: An Open Source Dynamic Simulator for Indoor Propagation and Localization

N. Amiot, M. Laaraiedh, B. Uguen PyLayers: An Open Source Dynamic Simulator for Indoor Propagation and Localization. In Proceedings of the IEEE International Conference on Communications 2013: IEEE ICC’13 - Workshop on Advances in Network Localization and Navigation (ANLN), Budapest, Hungary, June. 2013.. c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/ republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

58 / 146

PyLayers: An Open Source Dynamic Simulator for Indoor Propagation and Localization Nicolas Amiot, Mohamed Laaraiedh , Bernard Uguen University of Rennes 1, IETR Email: {nicolas.amiot, mohamed.laaraiedh, bernard.uguen}@univ-rennes1.fr Abstract—PyLayers is a new open source radio simulator. It has been designed to evaluate localization algorithm performances through the realistic simulation of location-dependent parameters (LDPs) in heterogeneous mobile radio networks. The radio channel is synthesized by using a novel graph-based ray tracing method which has been introduced in order to improve performances in mobile ray-tracing scenarios where geometrical information reuse from one mechanical time-step to another is advantageous. PyLayers can synthesized the narrow band, wide band or ultra wide band (UWB) channel impulse response and thus allows to produced various kind of location dependent parameters as the widely used LDPs received power an time of arrival. Realistic movement of pedestrian agents into the building layout is modeled with a virtual forces approach. The simulated data can be directly exploited with one of the original builtin localization algorithms or be exported to various standards file extensions for external post-processing. Examples of typical PyLayers outputs are provided . Keywords—Simulator, propagation, localization, ray-tracing, multi-wall, mobility model, graph theory

I. I NTRODUCTION VALUATING performances of an indoor localization algorithm is often a complicate task because performances highly depend on the chosen test environment. Three main approaches are generally retained to perform such an evaluation. First one, consists in using statistical models which are very efficient tools for obtaining a global trend of the algorithm accuracy. However, when the simulation scenario becomes very specific, this approach is often not sufficient. For more complex scenarios, a measurement campaign can be realized. An intermediate solution is to use simulation which can be rather accurate compare to measurements and sometimes more useful than statistical models in localization context. Nevertheless, the accuracy of the simulated data are conditioned by the realism level of the simulator. Most of available radio simulator such as [1] are not especially adapted for localization purposes. Those tools are generally built for specific bandwidths for a given standard, and this prevent the validation of heterogeneous localization algorithms. Furthermore, movements of persons are generally simulated with random models or more generally by a Markov chain model [2]. However, those approaches only capture the human mobility at a large scale, and don’t take into consideration small scale as attracting points of a room. The aim of PyLayers [3] is to provide a transversal, versatile and upgradeable simulator for indoor propagation and

E

localization: Transversal because it addresses multiple simulation issues (mechanical, radio, localization); Versatile because each part of simulation is independent and different type of radio access technologies (RATs) are envisaged; Upgradeable because it is an open source project, coded with Python and standard scientific high level libraries Numpy, Scipy, NetworkX and Matplotlib. In the following, the four specific fields of simulation, so-called layers, of PyLayers are presented from section II-A to section II-D. Then, the simulator outputs are presented in III. II.

P Y L AYERS STRUCTURE : A MULTI -L AYERS DESCRIPTION

PyLayers is a dynamic simulator for indoor propagation and localization based on graphs description. Fig.1 illustrates the four layers of the tool: layout, mechanical, network and localization. Each layer stands for a specific behavior of the simulator and operates independently. Data can be shared between layers, e.g. the positions computed at the mechanical layer can be used by the network layer. PyLayers

Localization

Network

Mechanical

Layout

Fig. 1: Conception of PyLayers in four independent layers.

A. The Layout Layer The layout layer uses an original graph description to describe the simulation scene, in terms of outlines, identification of rays and mobility of pedestrian agent. The five defined graphs are:

• The structure graph Gs (Fig.2) • The topological graph Gt (Fig.3) • The graph of rooms Gr (Fig.4) • The visibility graph Gv (Fig.5) • The interaction graph Gi (Fig.6) The structure graph Gs is used to describe the layout outlines and the materials used in the scene. Conventionally, the nodes with a positive index describes segments, whereas the nodes with a negative index describe points. In addition, segments have both a property attribute which indicate the slab reference or the segment height, and a transition attribute to indicate if the segment is traversable. Finally, the edges of Gs link two nodes with a positive and a negative index respectively. The topological graph Gt establishes the adjacency between cycles of Gs . Its nodes are the cycles of Gs , and its edges link adjacent cycles. The graph of rooms Gr is a subgraph of Gt which retains only cycles with at least one transition. It is used by the mechanical layer, to determine paths between rooms. The visibility graph Gv shares its nodes with Gs . Edges of Gv link nodes on an optical visibility criteria. Derived from Gv , the interaction graph Gi describes all possible interactions between nodes of Gv . Hence, on each node, up to 3 Gi nodes can appear in regard to the 3 types of available interactions (transmission, reflexion and diffraction). Nodes are specific tuples for each interaction type: transmission (Gs node id, in-room id, outroom id), reflection (Gs node id, room id), diffraction (Gs node id). This graph is used to perform the ray-tracing.

2

0

Fig. 4: The graph of room Gr . 13

13

14

-7

6

-8

17

-14 4 -15

18

7

3

-5

18

-6 9

5

3

-5 8 1

2

Fig. 5: The visibility graph Gv . (13, 2)

(12, 2)

(5, 0)

(1, 0)

9

5

-14 4 -15

7

-9

-6

17

6

-1

12

14

12

(6, 0) -3

1

(14, 2)

(4, 2, 1) (18, 1, 2) (17, 1, 2) (17, 1) (4, 2) (18, 1) (17, 2) (4, 1) (18, 2) (7, 0) (17, 2, 1) (4, 1, 2) (18, 2, 1) (7, 0, 1) (7, 1, 0) (7, 1) (9, 0) (9, 0, 1) (3, 1) (9, 1, 0) (9, 1) (8, 0) (8, 0, 1) (8, 1, 0) (8, 1)

(2, 1)

Fig. 6: the interaction graph Gi .

8 -2

1

-11

2

-10

Fig. 2: the structure graph Gs .

2

0

1

Fig. 3: The topological graph Gt .

B. The Mechanical Layer To ensure dynamic and layer independence, all the upcoming layers processing are based on the SimPy library [4]. SimPy is a discrete event simulator which allows the simulation of active components. In particular, the mechanical

layer uses the Personal Rapid Transit Simulation [5], a SimPybased library which defines the movement of pedestrian agents. Agents are independent instances moving into the layout. Both a large scale and a small scale description of the mobility are considered in PyLayers. The large scale description of the mobility helps to define where and how the agents goes. Practically, it consists in: choosing a target and find a path to reach that target. The target is a room and the path is a succession of rooms. The path from a room i to a targeted room t is obtained with the help of graph Gr . Rooms i and t corresponds respectively to nodes vri and vrt of Gr . Then, the path is obtained with a Djikstra [6] shortest path algorithm D: Vp = D(vri , vrt ).

(1)

The ordered set of node Vp describes the shortest path between vri and vrt . To achieve the whole trajectory from i to t, the agent go iteratively through all the rooms corresponding to the nodes of Vp . Each time a node of Vp is reached, next one becomes an intermediate target it, until the node vrt is reached. In Fig. 4, an agent in room i = 0 with a target in room t = 2, corresponds

to returned path Vp = {0, 1, 2}. Thus, the agent will first move from room 0 to its intermediate target 1. Once the agent has reached the center of room 1, it moves to its final target in room 2. Once their target are known and their path have been computed, agents can start moving in layout environment, avoiding walls and going through doors. This ability is defined as the small scale description of the mobility and is modeled using magnetic forces [7]. The magnetic forces model uses a resulting acceleration vector a to drive the agents through the layout. The magnetic force model supposes that the agents and the environment layout are modeled by positive poles, while the intermediates target rooms are modeled by negative poles. Each agent k is potentially under the influence of several magnetic forces: an attractive force Fit which pull the agent to its intermediate target node vit ; several repulsive forces Fr , which push the agent off the walls or other agents. According to [7], it is possible to approximate the attractive magnetic force Fkit on agent k, with the Coulomb’s law: qk qit ˆ Fkit ∝ Rk,it , (2) ||Rk,it ||2

with qk and qit being the intensities of the magnetic load of the agent and the intermediate target, and Rk,it being the vector from the agent k to the intermediate targeted node vit . The repulsive magnetic force Fr which avoids the agent penetrating a wall is given by: α ˆ k,w , d (3) Fr (dk,w ) = ||dk,w ||2

with dk,w the distance vector from the agent k to a wall w and α a parameter to adjust the repulsion, and thus avoiding the agent grazing the wall. Similarly, the inter-agent avoidance can be modeled with (3), by replacing dk,w by dk,l where k and l refers to two different agents. Then, the agent k resulting acceleration vector ak is written as the sum of the three previously described interactions: ak = Fkit +

W X w

Fr (dk,w ) +

K X

Fr (dk,l ),

(4)

l6=k

where W is the number of walls in the vicinity of the agent k, which is determined with a simple distance threshold. The use of that threshold avoids the computation of forces from all the walls of the entire layout. An example of operation of the mechanical layer can be seen in [8]. C. The Network Layer At the network layer all the possible radio links are determined at each time step. The network layer is defined as a graph Gn , where the nodes Vn of network graph have several attributes as type (mobile agent or anchor), emitted power in dBm, receiver sensitivity in dBm, ... In order to manage multiple RATs, Gn is built as a multigraph [9] where several parallel edges can link two nodes as node 1 and 2 on Fig 7. Each RAT is represented by a different edge of the graph. On the edges of Gn , are computed the LDPs of the link.

7

2

6

1 ra t2 ra t1

Fig. 7: Example of graph of network Gn which shows all the links between 4 radio nodes on 2 different RATs.

In order to get a fast determination of the network connectivity it is desirable to have an efficient approach especially in dynamic context. This task is realized thanks to an enhanced COST 231 model which can provide a fast indication of the connectivity based on predefined thresholds and analytic determination of packet error rate based on formulas described in [10]. The adopted path loss determination approach takes into account the slab thickness and constitutive material information as a function of angle and frequency. It allows to determine pretty accurately the amount of energy around the first path and the excess wall crossing time of flight which could be exploited as an UWB LDPs. Alternatively, the LDPs can be obtained by a ray-tracing approach from channel impulse response (CIR) computation. This approach requires a prior graph-based step: the determination of signatures. These signatures are then used to obtain rays. Let T be a device which tries to obtain LDPs from R, another device. The first step of signature computation consists in obtaining the two sets of nodes VT and VR which are in direct visibility of T and R respectively. Then, a shortest path algorithm is computed on Gv for each pair of nodes vT and vR of the set VT and VR respectively. Each obtained minimal sets of nodes between T and R is the so-called signature S. The procedure is described in Algorithm 1. Algorithm 1 determination of signatures Require: T ,R Require: Gs ,Gr ,Gv L = ∅ Initialize a list of Signatures VT ← get visible nodes(Gs , T ) VR ← get visible nodes(Gs , R) for nt ∈ VT do for nr ∈ VR do S = all-simple-paths(Gv , nt , nr ) L ← L.append(S) end for end for Because only optical visibility is taken into account during the determination of signatures, each signature is not necessarily valid to produce a ray. For instance, a signature may contains a transmission through a metallic segment. Hence, the validity of each signature is tested with the help of Gi . 2D rays are thus obtained from valid signatures. Finally, Those rays are expanded in 3D using simple geometric transformations. A complete description of this process is described in [11].

SIMULATOR OUTPUTS

Ma p of re c e ive d powe r g ra y : Pr (dBm ) < -80.00 dBm Pt = 0 dBm white : Pr (dBm ) < -95.84 dBm

30 36 48 54

dBm

42

60 66 72 78

Fig. 8: Map of received power. The black triangle is the transmitter position. The white area represents power under the noise floor, the grey area represents power under receiver sensitivity, elsewhere is the received power in dBm

In the following, several example of PyLayers outputs are illustrated. As explained before, LDPs can be obtained either from the enhanced COST 231 model or by the graph based raytracing tool. An example of full coverage of the received power along the direct path and the extra time of flight delay obtained with the enhanced multi-wall model is presented on Fig. 8 and Fig. 9 respectively. In addition of the received power value, information on areas under the receiver sensitivity and areas under the noise floor is also displayed. For the presented figure, the full coverage of 80 × 40 links takes only few seconds.

4.8 4.2 3.6 3.0 2.4 1.8 1.2 0.6 0.0

Fig. 9: Map of LOS excess delay. The white triangle is the transmitter position. The degraded colors represent the delay in ns introduced by the walls in addition of the LOS delay.

Notice that on Fig. 9 the high values of delays corresponds to region where there is no radio connectivity. This excess time of flight can be exploited in algorithm to build a ToA bias model. A more realistic model can be obtained from the ray-tracing CIR. Volta g e (mV)

III.

PyLayers provides numerous outputs and visualizations for data post-processing and localization analysis. From each layer, relevant information can be exported in standard format file (csv, Matlab). PyLayers exports for each agent timestamped position, velocity and acceleration vectors, LDPs and estimated positions.

Ma p of LOS e xc e s s de la y, pola r orthog ona l

8.0 6.0 4.0 2.0 0 2.0 4.0 6.0

Volta g e (mV)

D. The Localization Layer The localization layer consists of a set of specialized algorithmic modules which can be invoked on a distributed mode by each agent. Those modules can also be used on a centralized mode in post-processing the produced simulated data of each node, hence emulating a centralized processing node. The algorithmic modules are split in two distinct categories of methods: algebraic based method [12] and geometric based methods [13]. Algebraic based methods combine prior knowledge of anchor nodes position obtained either from side information or earlier estimation, and current available LDPs coming up from the running network layer. This set of methods includes classical algorithms often taken as reference in performances comparison as basic least square or weighted least square as well as more elaborated approaches based on closed-form expressions for specialized likelihood functions on various heterogeneous and hybrid (different LDPs) situations. Geometric based methods exploit any LDPs or any piece of available prior information as a geometrical constraint which implicitly introduces a given partitioning of subset of 3D space. PyLayers provides a computational implementation of an original set membership techniques which has shown being robust on real measurements [14] and compare very well with concurrent algebraic approaches.

8.0 6.0 4.0 2.0 0 2.0 4.0 6.0

Me a s ure m e nts

0

25

Tim e (ns )

50

Simulations

0

25

Tim e (ns )

50

Fig. 10: Example of measured and simulated CIRs.

An example of synthesized CIR from the graph based raytracing tool is shown and compared to measurements [15] in Fig. 10. The purpose here is not to demonstrate a perfect fitting with measurement which remains a deceptive goal with respect to the lack of ground truth information present in such comparison. The interest of ray-tracing tool synthesis in the localization context is precisely that it can predict quite well those paths due to the intrinsic building geometry which are precisely this information which is going to be relevant to exploit in a multi-path localization scheme [16]. The previously shown high level propagation functionalities allows PyLayers to deal with realistic localization scenario putting together highly site specific dependent LDPs and indoor pedestrian mobility. As an illustration, Fig.11 presents two localization outputs of PyLayers. The figure on the top shows the result of dynamic scenario where an agent is moving into a layout and estimates its position with the help of fixes anchors. The 500 green circles are the different positions of the agent, and the red circles are the anchor positions. Below, a cumulative density function compares the performances of the two proposed categories of localization algorithm for the given scenario. Curves shows the advantage of the geometric based on a weighted least square for 85% of estimated positions.

1.0

Cumulative Probability

Geometric Algebraic 0.8

0.6

0.4

0.2

0.0

0

2

4

6

8

10

Positioning error [m]

Fig. 11: CDF comparison between geometric and algebraic weighted least square method, on a simulated synthetic scenario. The considered scenario is shown on the top. Red circles represents the anchors nodes and green circles represents the 500 positions of the agent used for plotting the CDF.

IV. C ONCLUSIONS This paper has presented an overview of PyLayers, an open source dynamic radio simulator for indoor propagation and localization. The simulator exploits a multi-graph description of the layout. This description allows to provide an accurate human-like mobility model, an enhanced multi-wall model, and a novel graph-based ray-tracing tool. In addition, algebraic based algorithm and original geometric based algorithm are provided. Visualizations of outputs from the simulator have been shown. A comparison against measurements has shown that the simulator is able to provide accurate CIRs and LDPs. Our current work consists in both improving the global computation speed, and add cooperative localization scheme. Future work consists in implementing MAC layer mechanisms. Finally, due to the open source aspect of the project, we also invite any interested person to use, test, or propose improvements of any parts of the project, by forking the following GitHub repository : https://github.com/PyLayers/PyLayers. ACKNOWLEDGMENT The work presented in this paper has been performed in the framework of the FP7 project ICT-248894 WHERE2 (Wireless Hybrid Enhanced Mobile Radio Estimators - Phase 2) which is funded by the European Union. R EFERENCES [1]

F. Esparza, V. Torres, M. Beruete, A. Lopez, and F. Falcone, “Simulation of indoor lte behaviour,” in Antennas and Propagation (EuCAP), 2010 Proceedings of the Fourth European Conference on, pp. 1–3, april 2010. [2] J. Nielsen, R. Olsen, T. Madsen, and H. Schwefel, “On the impact of information delay on location-based relaying: A markov modeling approach,” in Wireless Communications and Networking Conference (WCNC), 2012 IEEE, pp. 3045–3050, april 2012.

[3] Pylayers available at http://www.pylayers.org. [4] SimPy available at http://simpy.sourceforge.net/. [5] K. MacLeod, Personal Rapid Transit Simulation, available at : http://sourceforge.net/projects/prt/. [6] E. Dijkstra, A Short Introduction to the Art of Programming. Holland, 1971. [7] K. Teknomo, Microscopic Pedestrian Flow Characteristics: Development of an Image Processing Data Collection and Simulation Model. PhD thesis, Tohoku University, Japan, Sendai, 2002. [8] PyLayer mechanical layer example : http://goo.gl/hCFtR. [9] B. Bollob´as, Modern Graph Theory: B´ela Bollob´as. Graduate Texts in Mathematics Series, Springer, 1998. [10] M. Zuniga and B. Krishnamachari, “Analyzing the transitional region in low power wireless links,” in In First IEEE International Conference on Sensor and Ad hoc Communications and Networks (SECON, pp. 517– 526, 2004. [11] M. Laaraiedh, N. Amiot, and B. Uguen, “Efficient ray tracing tool for UWB propagation and localization modeling,” in 2013 7th European Conference on Antennas and Propagation (EuCAP) (EuCAP 2013), (Gothenburg, Sweden), Apr. 2013. [12] M. Laaraiedh, L. Yu, S. Avrillon, and B. Uguen, “Comparison of hybrid localization schemes using rssi, toa, and tdoa,” Wireless Conference 2011 - Sustainable Wireless Technologies (European Wireless), 11th European, pp. 1–5, april 2011. [13] N. Amiot, M. Laaraiedh, and B. Uguen, “Evaluation of a geometric positioning algorithm for hybrid wireless networks,” in Software, Telecommunications and Computer Networks (SoftCOM), 2012 20th International Conference on, pp. 1–5, sept. 2012. [14] B. Denis, R. Raulefs, B. H. Fleury, B. Uguen, N. Amiot, L. De Celis, J. Dominguez, M. Koldsgaard, M. Laaraiedh, H. Noureddine, E. Staudinger, and G. Steinboeck, “Cooperative and heterogeneous indoor localization experiments,” in IEEE International Conference on Communications 2013: IEEE ICC’13 - Workshop on Advances in Network Localization and Navigation (ANLN) (ICC’13 - IEEE ICC’13 - Workshop ANLN), (Budapest, Hungary), June 2013. [15] ICT-WHERE-Project, “Deliverable 4.1: Measurements of locationdependent channel features,” tech. rep., October 2008. [16] K. Pahlavan, F. Akgul, M. Heidari, A. Hatami, J. Elwell, and R. Tingley, “Indoor geolocation in the absence of direct path,” Wireless Communications, IEEE, vol. 13, no. 6, pp. 50–58, Dec. 06.

ICT–248894 WHERE2

A.4

D2.6

Efficient Ray Tracing Tool for UWB Propagation and Localization Modeling

M. Laaraiedh, N. Amiot, B. Uguen Efficient Ray Tracing Tool for UWB Propagation and Localization Modeling. In Proceedings of the 7th European Conference on Antennas and Propagation (EuCAP 2013), Gothenburg, Sweden, Apr. 2013.. c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/ republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

64 / 146

Efficient Ray Tracing Tool for UWB Propagation and Localization Modeling Mohamed Laaraiedh, Nicolas Amiot, Bernard Uguen University of Rennes 1, IETR Email: [email protected]

Abstract—Ray tracing simulators show higher accuracy in modeling propagation of radio waves. This RT based modeling has been used for designing wireless communications systems. Recently, these RT tools have been applied for the designing of localization techniques and systems. This paper regards the application of ray tracing tools in radio propagation modeling for ultra wide band localization applications. A graph-based ray tracing technique is presented and used to simulate an indoor environment. The obtained simulations are then post processed in order to extract location-dependent parameters such as RSSI and TOA and compare them to real measurements carried out in the same environment. The outcome of this comparison is a good match between simulations and measurements even without considering pieces of furniture presented in the environment during the measurement campaign.

I. I NTRODUCTION Ray tracing (RT) tools have been developed for accurate radio propagation predictions [1], [2], [3], [4], [5]. Those RT tools are based on geometrical optics (GO) and geometric/uniform theory of diffraction (GTD/UTD) which allow to define the optical paths between two points using an electromagnetic formulation of propagation phenomena such as reflection, transmission, and diffraction. In order to perform simulation with RT tools, a detailed 3D site geometry, material properties, transmitter/receiver locations, and antenna properties are required. Once this complete information is given, a RT task is executed. Then, after the identification of rays that are captured by the receiver antenna, the channel transfer function is calculated. Conventional RT tools takes usually much time to simulate a radio link. This delay makes those tools not suitable for real time communication and localization techniques. In this paper, a new technique of RT is presented. This technique is based on graph theory i.e the modeling of environment layouts using graphs. A faster technique to compute rays is proposed based on those graphs [6]. The proposed technique is applied to estimate location dependent parameters (RSSI and TOA). The comparison between measurements and simulations reveals a good match even without taking into account the pieces of furniture present in the environment during measurements. These results would allow to use RT tools in localization systems to complement/replace the measurement campaigns. II. P Y L AYERS :

GRAPH - BASED RAY TRACING SIMULATOR

A. Graph-based modeling of indoor environments PyLayers simulator [7] describes the data structure of the indoor layout through the definition of the following graphs :

The structure graph Gs The visibility graph Gv • The topological graph Gt • The graph of rooms Gr • The graph of interactions Gi Definitions and examples of these graphs are given in Figures 1-5. The adopted multi-graph description contains meta information from the layout which can be exploited for both incremental identification of rays and simulation of indoor pedestrian mobility. • •

B. Ray tracing Let first define a ray with Ni interactions as an ordered list of Ni + 2 points starting at transmitter position tx and ending at receiver position rx . R = {tx , p0 , · · · , pNi −1 , rx } = {tx , P, rx }

(1)

S(R) = S(P) = {Id(p0 ), · · · , Id(pNi −1 )}

(2)

where P is the ordered list of Ni interaction points {p0 , · · · , pNi −1 }. The ray tracing (RT) consists then in determining the list P. Based on the graphs described previously, we first construct rays signatures and then we apply a twosteps technique to obtain 2D rays from signatures. Those 2D rays are then transformed into 3D rays using geometric transformations and theorems (i.g. Thales). 1) ray signature: The signature of R is defined as the sequence of structure ids of P. The structure id of a point is simply the identification number either of the diffraction point (id < 0) or the segment number (id > 0), this is directly a node number from graph Gs or Gv defined previously. Id(p) is a function which returns the layout index number of p.

Conventionally, the signature of the LOS path is the empty set S({tx , rx }) = ∅. Notice that when at least one ending point of the radio link is moving, all the points of P are moving correspondingly. This is not the case of the signature P which can remain stationary in the vicinity region of the moving point. In a first step the algorithm seeks for the nodes Vtx of Gs which are visible from tx, and the set of nodes Vrx of Gs which are visible from rx. The second step consists in exploiting the 2 subsets Vtx and Vrx for determining Nt × Nr first order signatures where card(Vtx ) = Nt and card(Vrx ) = Nr . The algorithm determines then all the simple paths (paths with no repeated nodes) between any pair of nodes in the graph Gi .

-6

5

6 -1

1

-5 7 -7 9 -8 8 -2

4

-4 3

2

-3

Fig. 1: Gs representation of a 2 rooms separated by a door (in red), positive nodes are walls and negative nodes are vertices. 5

4 7 -7 9 -8 8

6 1

3 2

Fig. 2: Visibility graph Gv : has the same nodes of Gs but edges are defined differently. Two nodes are connected if they share an optical visibility relationship.

0

1

Fig. 3: Topological graph Gt . It results from the determination of all cycles of the graph Gs . Gt contains a valuable topological information. Each cycle contains various attributes as for example the set of nodes and edges of the cycle.

0

1

Require: tx ,rx Require: Gs ,Gv ,Gr L = ∅ Initialize a list of signatures Vtx ← get visible nodes(Gr , tx ) Vrx ← get visible nodes(Gr , rx ) for nt ∈ Vtx do for nr ∈ Vrx do Sit,ir = all-simple-paths(Gi , nt , nr ) L ← L.append(Sit,ir ) end for end for

2) from signature to ray: The determination of a ray from its signature is a two steps process. The fist step consists in calculating a set of intermediate points starting from the transmitter tx = [xt , yt ]T . Once calculated those intermediate points can be used for any receiver position with the same signature. The signature can be either valid or not valid for a given receiver depending on its coordinates. We consider a signature of length N which can contain three types of interaction: diffraction (D), reflection (R), transmission (T). From this signature, 2 ordered lists of points are retrieved. pa = [a0 , ..., aN −1 ]

(3)

pb = [b0 , ..., bN −1 ]

(4)

ak = [xAk , yAk ]T , bk = [xBk , yBk ]T

(5)

In the case of a diffraction the point appears identical in each list otherwise pa and pb contains respectively tail points and head points of the signature segments. The sequence of intermediate points can be determined from the knowledge of the transmitter point, the sequence of segments and also importantly the type of interaction (R|T |D). This sequence of intermediate points is: m = [m0 , ..., mN −1 ]

Fig. 4: graph of rooms Gr . It is the subgraph of Gt retaining only cycles which possess at least one door (i.e. doors). (5, 0)

(4, 1) (7,71) (7, 0) -7 (9,91) (9, 0) -8

(6, 0)

(3, 1)

(8,80) (8, 1) (1, 0)

(2, 1)

Fig. 5: graph of interactions Gi . It results from the determination of all possible interactions between the nodes of Gs . The nodes of Gi are either a tuple (id of Gs node, id of room) in case of reflection or a simple number (negative for diffraction and positive for transmission)

(6)

with mk = [xk , yk ]T . Depending on the very nature of the interaction the intermediate point is either • (R) : the image of the current point w.r.t the segment, • (T) : the current point itself • (D) : the diffraction (interaction) point itself m0 depends on the type of the first interaction as: • (R) : m0 = S0 tx + v0 • (T) : m0 = tx • (D) : m0 = a0 and similarly, mk depends on the type of the k th interaction: • (R) : mk = Sk mk−1 + vk • (T) : mk = mk−1 • (D) : mk = ak where Sk (reflection matrix) and vk are defined as:   ak −bk Sk = (7) −bk −ak

The translation vector is defined as : vk = [ck , dk ]T

(8)

with (xAk −xBk )2 −(yAk −yBk )2 (xAk −xBk )2 +(yAk −yBk )2 2(xBk −xAk )(yAk −yBk ) (xAk −xBk )2 +(yAk −yBk )2

ak = bk = ck =

(9)

Bk ) +yAk (xBk −xAk )(yAk −yBk ) 2 xAk (yAk −y (xAk −xBk )2 +(yAk −yBk )2 2

Bk )(xBk −xAk )+yAk (xBk −xAk ) dk = 2 xAk (yAk −y (xAk −xBk )2 +(yAk −yBk )2

(12)



  y= 

I2 .. .

   

..

. {−SN −1 | − I2 |O2 } I2 (14)  {S0 tx + v0 |tx |a0 }  {v1 |z2 |d1 }  (15)  ..  .

C. Determination of interaction points Let the receiver be noted by rx = [xr , yr ]T . Interaction points are points which belong to the layout segments. They are obtained from the intersection between the segments indicated by the signature and the segments joining successively the previously obtained intermediate points m. The unknown 0 0 interaction points are pk = [xk , yk ]T . By introducing a parametrization on  T both the unknown 4T× 1 column vector 0 with Γk = [αk , βk ] . Starting from the is xk = pTk , ΓTk receiver and joining the last intermediate point mN −1 , we obtain: (16) (17)

p0 + α0 (rx − mN −1 ) = rx p0 + β0 (aN −1 − bN −1 ) = aN −1 We solve successively the following linear systems 0

0

(18)

T0 x0 = y0

T0 =



I2 I2

0



I2 I2

(21) (22) (23) (24)

 pk−1 − mN −(k+1) 02×1 02×1 aN −(k+1) − bN −(k+1) (25) iT h 0 T T (26) yk = pk−1 , aN −(k+1)

The parametrization of the signature segment plays a very important role, if 0 < βk , αk < 1 the interaction point pk is valid and the calculation can proceed otherwise the interaction point is out of the segment and the signature is rejected for the current receiver coordinates. An example of obtained 2D rays is plotted in Figure 6.

Tx Rx

{vN−1 |z2 |dN −1 }

where I2 and O2 are respectively identity matrix and zero matrix of dimension 2, z2 is 2 × 1 zero column vector. The notation {R|T |D} stands for the possible matrix entries depending on the interaction type.

with :

with : Tk =

(13) 

Am = y

0

Tk xk = yk

(11)

The coordinates of intermediates points are obtained by solving the following linear system:

I2  {−S1 | − I2 |O2 }  A= 

pk + αk (pk−1 − mN −(k+1) ) = pk−1 pk + βk (aN −(k+1) − bN −(k+1) ) = aN −(k+1)

(10)

2



The interaction point pk is determined identically taking advantage of the knowledge of point pk−1 :

rx − mN −1 02×1 02×1 aN −1 − bN −1  T 0 y0 = rx T , aTN −1



(19) (20)

Fig. 6: Example of obtained 2D rays. III. A PPLICATION OF RAY TRACING TECHNIQUE FOR THE ESTIMATION OF LOCATION - DEPENDENT PARAMETERS modeling of LDPs is very useful in evaluating localization techniques performances. These models should be as accurate as possible. Since RT tools are able to reproduce radio channel and propagation phenomena with higher accuracy, models extracted from RT simulations would be rather accurate for realistic evaluation of localization techniques [5]. In order to assess the RT tool accuracy in LDPs modeling, we compare the RSSI and TOA values obtained using the Pylayers simulator and those obtained during an UWB measurement campaign. Measurements and simulations are carried out in same scenario described in Figure.7 where 4 receivers (Rx1-4) are fixed and one transmitter (Tx) is moving on 302 positions into an indoor office environment [8]. For each position of the Tx, both a simulated and measured channel responses to an UWB pulse are recorded and post-processed in order to extract RSSI and TOA. Examples of simulated and measured channel responses are given in Figure 8. In Figure. 9 we plot for each Rx the evolution of measured and simulated RSSIs along the trajectory defined during the measurement campaign. Although the general patterns of measured and simulated RSSI look the same, some large differences occurs at some points. This is particularly explained by

Rx4

Rx2 Tx145

Rx3

Rx1

Fig. 7: Rx and Tx locations during the WHERE measurement campaign. Measurements

0.008

IV. C ONCLUSION

0.006

Voltage (V)

0.004 0.002 0.000 0.002 0.004 0.006

0

50

100 Time (ns)

150

200

Simulations

0.008

tion by the measurement process when calculating the CIR unlike in simulations where the environment description is limited to walls, doors and windows. As an outcome from this comparison, the reliability of a RT tool in LDPs modeling depends on the precision and the depth of environment description. A complete environment description in which walls, doors, windows, and furniture are considered and properly described would lead to better simulation accuracy and avoid hence carrying laborious measurements.

0.006

In this paper we have presented a graph-based ray tracing tool. The proposed technique exploited in radio propagation modeling for localization techniques. Investigations and measurements-simulations comparison showed that ray tracing tools can fairly estimate location dependent parameters (RSSI and TOA) once a proper description of propagation environments and antennas is performed. Ray tracing tools represent a good alternative to measurement campaigns which are usually very laborious and painful to carry out. Unlike measurements, the ray tracing tools offer also the possibility of updating the description of environments and antennas according to the studied scenario. Next step will be to include pieces of furniture (especially metallic pieces) and re-evaluate the performance.

Voltage (V)

0.004

ACKNOWLEDGMENT

0.002

This work was performed in the framework of the WHERE2 (ICT-248894) project, which is partly funded by the European Commission.

0.000 0.002 0.004 0.006

R EFERENCES 0

50

100 Time (ns)

150

200

Fig. 8: Simulated vs. measured channel response for a given link (Rx2 and Tx145 in Figure 7).

the fact that many pieces of furniture have not been included in the environment description. In particular, metallic cabinets (small rectangles in Figure.7) have strong effects on radiopropagation and thus on the RSSI values. In order to reduce the gap between simulations and measurements, these pieces of furniture must be properly included in the description of the environment. As for TOA, we apply on both simulated and measured CR the Thresholded dichotomous left interval selection technique put forward in [9]. In the same way as for RSSI, we plot in Figure.10 the evolution of estimated ranging errors along the same trajectory for both simulations and measurements. These figures show that the regions with large RSSI simulation errors (in Figure.9) coincide with the regions where the ranging error obtained using measurements is very large. This is explained by the presence of metallic cabinets which obstruct the direct path. These metallic cabinets are in fact taken into considera-

[1] J. McKown and J. Hamilton, R.L., “Ray tracing as a design tool for radio networks,” Network, IEEE, vol. 5, pp. 27 –30, nov. 1991. [2] S. Seidel and T. Rappaport, “Site-specific propagation prediction for wireless in-building personal communication system design,” Vehicular Technology, IEEE Transactions on, vol. 43, pp. 879 –891, nov 1994. [3] P. Meissner, D. Arnitz, T. Gigl, and K. Witrisal, “Analysis of an indoor UWB channel for multipath-aided localization,” in Ultra-Wideband (ICUWB), 2011 IEEE International Conference on, pp. 565 –569, sept. 2011. [4] M. Raspopoulos, C. Laoudias, L. Kanaris, A. Kokkinis, C. G. Panayiotou, and S. Stavrou, “3D Ray Tracing for Device-Independent Fingerprintbased Positioning in WLANs,” in 2012 Workshop on Positioning, Navigation, and Communication (WPNC), March 2012. [5] M. Laaraiedh, B. Uguen, J. Stephan, Y. Corre, Y. Lostanlen, M. Raspopoulos, and S. Stavrou, “Ray tracing-based radio propagation modeling for indoor localization purposes,” in Computer Aided Modeling and Design of Communication Links and Networks (CAMAD), 2012 IEEE 17th International Workshop on, pp. 276 –280, sept. 2012. [6] B. Uguen, N. Amiot, and M. Laaraiedh, “Exploiting the Graph Description of Indoor Layout for Ray Persistency Modeling in Moving Channel,” in Proceedings of the 6th European Conference on Antennas and Propagation (EuCAP 2012), (Prague, Czech Republic), March 2012. [7] PyLayers simulator, “http://www.pylayers.org/.” [8] Y. Lostanlen, J. Stephan, J. Keignart, W. Wang, D. Slock, and F. Kaltenberger, “D4.1: Measurements of location-dependent channel features.” Deliverable, October 2008. [9] L. Yu, M. Laaraiedh, S. Avrillon, B. Uguen, J. Keignart, and J. Stephan, “Performance Evaluation of Threshold-Based TOA Estimation Techniques Using IR-UWB Indoor Measurements,” in Proc. of 18th European Wireless Conference (EW2012), (Poznan, Poland), April 2012.

−40

−40

Rx1

−60 −70 −80 −90

Measurements RT Simulations

−50

RSSI (dB)

RSSI (dB)

−50

Rx2

−60 −70 −80

0

50

100

150

200

250

−90

300

0

50

Trajectory Index −40

−40

Rx3

200

250

300

250

300

250

300

250

300

Rx4

−50

RSSI (dB)

RSSI (dB)

150

Trajectory Index

−50 −60 −70 −80 −90

100

−60 −70 −80

0

50

100

150

200

250

300

−90

0

50

Trajectory Index

100

150

200

Trajectory Index

Fig. 9: Measured versus Simulated RSSI values for different receivers. 10

10

Rx1

8

Ranging error (m)

Ranging error (m)

8

6

4

2

0

Rx2

Measurements RT Simulations

6

4

2

0

50

100

150

200

250

0

300

0

50

Trajectory Index 10

10

Rx3

200

Rx4

8

Ranging error (m)

Ranging error (m)

150

Trajectory Index

8

6

4

2

0

100

6

4

2

0

50

100

150

200

Trajectory Index

250

300

0

0

50

100

150

200

Trajectory Index

Fig. 10: Measured versus Simulated TOA ranging values for different receivers.

ICT–248894 WHERE2

A.5

D2.6

Ray Tracing-Based Radio Propagation Modeling for Indoor Localization Purposes

Mohamed Laaraiedh, Bernard Uguen, Julien Stephan, Yoann Corre, Yves Lostanlen, Marios Raspopoulos, Stavros Stavrou. 17th International Workshop on Computer-Aided Modeling Analysis and Design of Communication Links and Networks (CAMAD 2012), September 17-19, 2012, in Barcelona. c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

70 / 146

Ray Tracing-Based Radio Propagation Modeling for Indoor Localization Purposes Mohamed Laaraiedh† , Bernard Uguen† , Julien Stephan‡ , Yoann Corre‡ , Yves Lostanlen‡ , Marios Raspopoulos] , Stavros Stavrou∗ † Universit´ e de Rennes 1 - IETR, av. du G´en´eral Leclerc, 35042 Rennes, France ‡ SIRADEL, 3 all´ ee Adolphe Bobierre, 35043 Rennes, France ] Sigint Solutions Ltd, Nicosia, Cyprus ∗ Faculty of Pure and Applied Sciences, Open University of Cyprus [email protected], [email protected], [email protected], [email protected]

Abstract— This paper regards the application of ray tracing (RT) tools for indoor localization techniques. The paper presents two possible applications of RT tools in that context. The first is the train/construction of fingerprints databases in order to complement or/and replace laborious measurement campaigns. The second application is the modeling of location-dependent parameters in order to feed localization techniques. These two applications are investigated and validated using three RT tools applied on a synthetic indoor environment. The obtained results showed that RT tools are reliable in both fingerprinting databases construction and LDPs modeling. Index Terms— Ray tracing, Localization, Radio propagation modeling, Simulation, Ultra wide band, Wide band, Narrow band

I. I NTRODUCTION Radio propagation modeling is of great interest in optimizing communications and localization systems in order to offer to users the most reliable systems and services. In order to avoid, as far as possible, measurement campaigns which are usually painful to carry out and consume huge resources, ray tracing (RT) tools are a good alternative approach for the radio propagation modeling. RT tools play, since many years ago, a relevant role in the design of modern communication systems [1], [2]. Recently, these RT based modeling tools are also used to design localization systems especially in indoor environments [3], [4]. Future communications and localization systems are mainly characterized by the cohabitation of heterogeneous radio access technologies (RATs). RT tools must hence be able to simulate heterogeneous channel RATs including Ultra-Wide Band (UWB), Wide Band (WB) and Narrow Band (NB) channels. A reliable RT modeling must hence cope with the different characteristics of these bandwidths in order to perform a valid modeling of heterogeneous radio channels. In that vision, three different RT tools have been exploited within the framework of FP7 projects WHERE and WHERE2. In this paper, these RT tools are first technically presented. These RT tools are then used to simulate the same predefined synthetic indoor environment for respectively UWB, WB and

NB bandwidths in order to finally obtain channel impulse responses (CIRs). From these CIRs, different location dependent parameters (LDPs) are extracted and analyzed from the very localization perspective. II. A DVANCED R AY T RACING S OLUTIONS RT tools have been developed for accurate radio propagation predictions. Those RT tools are based on geometrical optics (GO) and geometric/uniform theory of diffraction (GTD/UTD) which allow to define the optical paths from the transmitter to a receiver using an electromagnetic formulation of propagation phenomena such as reflection, transmission, and diffraction. In order to perform simulation with RT tools, a detailed 3D site geometry, material properties, transmitter/receiver locations, and antenna properties are required. Once this complete information is given, a RT task is executed. Then, after the identification of rays that are captured by the receiver antenna, the channel transfer function is calculated. In some tools the tracing task is preceded by a ray launching step where rays are uniformly propagated from the transmitter. During the WHERE projects, three 3D RT tools are being used in order to model heterogeneous radio propagation channels for both communication and localization systems design. These RT tools are respectively PyRay, Volcano, and 3DTruEM. Each of them is currently enhanced in the frame of the WHERE2 project to address a specific need of radio channel propagation modeling for indoor localization purpose. • PyRay is mainly an UWB indoor RT tool developed in Python. In order to simulate UWB antennas, PyRay uses vector spherical harmonics (VSH) to compactly describe the antenna radiation pattern within the entire frequency band [5]. VSH expansion of antenna radiation function simplifies data storage, process and analysis. Hence, the amount of data necessary to describe antennas is drastically reduced, the vector radiation functions are quickly reconstructed for different directions, and antenna radiation pattern is properly reconstructed [5]. PyRay is reinforced by a graph-based description of indoor channels based on four types of graphs: the structure

•

graph, the visibility graph, the topological graph and the graph of rooms [6]. Based on this graph representation and using Dijkstra’s algorithm, the signature of a ray is determined as the sequence of interactions of the ray from the transmitter to the receiver. In addition, a twostep process is developed in order to determine rays from their signatures. This ray determining process is used together with the graph-based representation of the radio channel to make faster the simulation of radio propagation channels [6]. Volcano engine by SIRADEL implements advanced propagation models to address the simulation of radio-wave propagation in a deterministic way and in all types of outdoor/indoor environments. Volcano comprises 3 models: Rural Model, Urban Model and Indoor Model. Volcano models are suitable for all complex environments as well as for large frequency range (from 90MHz to 10GHz+) and multi-bandwidths. In the frame of the WHERE2 project, SIRADEL currently enhances his Volcano Indoor Multi-Path model by including new advanced propagation techniques to address the time-variant and non-stationary effects of the indoor human crowd activity in a multilink context. This model is based on an accurate 3D geometrical description of the considered scene (from Computer-Aided Design files) containing the building structure and the position of the partition walls, the floors and the ceilings, the windows, the doors and potentially the furniture. In order to get an assessment of the received power and of the complex CIR for a specific radio link, the model considers the direct-path losses occurring between the transmitter and receiver as well as the multiple reflections (GO), diffractions (UTD) and transmissions through the walls and floors encountered by the multiple paths. Main input parameters of the model are the frequency, antennas and material dielectric properties, as well as some tuning coefficients that allow for measurement-based calibration. The model is able to consider human activity in order to simulate continuous and correlated time variations in the channel properties (path-loss, frequency response, power angle profiles and Doppler shifts) of different static indoor radio links. The methodology is basically as follows: a realistic timevariant scenario defines continuous movement of persons in the prediction environment; static channel properties are predicted from a 3D RT deterministic approach; the successive locations of human bodies are simulated into correlated snapshots (from a sampling of the simulation time); then the impact of each radio link channel is obtained from determination of ray obstructions, taking into account the multi-path ray geometries and the precise location of human bodies at each snapshot. The method is an extension of the model proposed in [7] adapted to multi-link deterministic channel predictions and a larger number of person distribution cases (especially when nobody obstructs the direct path). Used conjointly with the multi-RATs capability already mastered, the Volcano indoor model is becoming a very complete and valuable tool to address every geolocation purposes indoors.

•

3DTruEM is mainly a NB indoor RT tool which was developed in .NET programming languages by Sigint Solutions. The 3D environment description can be defined by means of a built-in 3D-CAD application or by importing standard CAD files. It offers the ability to define the receiver and transmitter antenna characteristics from a wide range of standard antennas but also the flexibility to import a custom-made antenna by importing its 3D radiation pattern/characteristics including also its polarization characteristics. In addition to its sophisticated algorithm that significantly improves speed, 3DTruEM also offers the ability of running a distributed-parallel multi-threaded simulation on a cluster of computers. In the framework of WHERE2, 3DTruEM has been extended to UWB simulations using band-divided Ray Tracing Method without modifying its calculation engine [8]. This technique allows notably to incorporate the frequency dependence of the electrical parameters of building materials which in many cases might be significant. III. H ETEROGENEOUS S IMULATIONS OF INDOOR PROPAGATION CHANNELS

As aforementioned, a realistic synthetic test environment is built in the frame of the WHERE2 project in order to enable a fair performance assessment of the designed indoor geolocation techniques. This WHERE2 synthetic test environment is initially composed of measurements collected during the WHERE project [9] and a detailed 3D description of the measurement environment (i.e. the first floor of a recent office building [10]). Since the collected indoor measurements only cover a limited set of scenarios (in term of radio links), RT predictions appear as a valuable solution to complement measurements and provide a generic evaluation framework. Measurements are used to validate the RT predictions and ensure high possible reliability (through a calibration step). Then, as into the measurement set, the RT is used to get UWB, WB and NB simulations in a common scenario. The resulting data consist of CIRs simulated on a grid of 51 pseudo-Access Points (APs) x 363 Mobile Stations (MSs) disseminated in the indoor environment. The selected simulation grid has been defined to provide sufficiently representative geometrical and channel configurations. Figure 1 shows locations of the APs (red dots) and MSs (black dots). Height for both APs and MSs is set at 1.5m above the ground floor. A constant transmitted power is assumed over the whole frequency band for NB and WB simulations, respectively over 100MHz and 500MHzbandwidth around 3.5GHz, whereas a waveform conforms to the 802.15.4a standard and centered at 4.493GHz is considered for UWB simulations. Each radio link simulation is computed by considering an isotropic antenna for the APs and successively isotropic and sector antennas with different azimuths (from 0◦ [North] to 315◦ by a step of 45◦ ) for the MSs. A theoretical 45◦ aperture in the horizontal plane is considered for the sector antennas. Figure 2 shows simulation examples from three different radio links using different RT tools. Finally, measurements and predictions form an enriched CIRs database, from which different location/context-radio parameters such as Received Signal Strength Indicators (RSSIs),

multipath Times of Arrival (TOAs), (Average) Power delay Profiles ((A)PDPs), channel delay spread or mean excess delay can be extracted.

Fig. 1: The simulated synthetic environment.

Narrow-band simulation between AP1 and MS38 (3DTruEM)

Wide-band simulation between AP1 and MS50 (Volcano)

Ultra Wide-band simulation between AP1 and MS62 (PyRay)

Fig. 2: Example of ray-tracing CIR simulations for three radio links using three different RT tools.

IV. A PPLICATIONS OF RAY TRACING TOOLS IN LOCALIZATION TECHNIQUES

A. Application to fingerprinting techniques The basic idea behind fingerprinting (FP) is to store the pre-defined position-dependent signal information (e.g. RSSI, PDP, etc.), known as fingerprints, for the entire coverage area of a wireless system in a database and then try to match or correlate a measured signal to the ones stored in the database in order to estimate the position. Generating and maintaining the database is very important and it can be either done through an extensive measurement campaign or through radio propagation modeling techniques. Experimental measurements might lead to more accurate fingerprints, but this process might be very laborious and also the applicability of the measured radiomap is reduced if the wireless environment is changed. Therefore, RT becomes a more attractive and efficient way to collect

fingerprints. In the WHERE2 project various techniques are proposed which utilize RT for advanced FP positioning and also alternative techniques to overcome possible deployments impairments and uncertainties [11]. In addition to the significant time and effort required to train a database another limitation of FP is the fact that the heterogeneity of devices introduces a variable that degrades the positioning performance when the device to be positioned is different from the one that the original measurements have been performed in the learning/training phase. Therefore, we have studied the use of 3D RT to construct radiomaps for WLAN RSSI fingerprint-based positioning, in conjunction with calibration techniques to make the overall process deviceindependent [4]. We addressed both challenges by exploiting 3D RT-generated radiomaps and using linear data transformation to match the characteristics of various devices. We evaluated the efficiency of this approach in terms of the time spent to create the radiomap, the amount of data required to calibrate the radiomap for different devices and the positioning error which is compared against the case of using dedicated radiomaps collected with each device. Our performance evaluation indicates that only a small amount of device-specific data are required to reach the same level of positioning accuracy attained with a manually collected radiomap (about 60% less data collection effort). Thus, this approach is far less laborious compared to traditional radiomap construction. Moreover, the radiomap can be easily updated if the propagation environment changes in the future (e.g. APs are added or removed, furniture or other equipment is relocated, etc.) by running the RT simulator, instead of collecting the radiomap data from scratch. In Fig.3, we plot the CDF of absolute positioning error obtained by applying a neural networks-based fingerprinting technique respectively on measurements and RT simulations obtained using PyRay tool. The technique applies feedforward neural networks to learn the 302 MS positions using four RSSIs measured (respectively estimated) with four APs as shown in Fig.4 which plots the MSs and APs positions defined during the WHERE measurement campaign carried in the same environment described in section III. Fig.3 shows that RT simulations give similar fingerprinting accuracy to that obtained using measurements which proves that RT tools can be reliable in such localization techniques. Although RT is powerful technique to generate FP radiomaps, its accuracy is strongly subject to the accuracy of the input parameters, such as the geometrical (dimensions) and morphological description (electrical parameters of walls) of the buildings and also other uncertainties such as the antenna pattern, location of clutter etc. It is common practice to use standard electrical parameters [12] for the various building materials to morphologically describe the environment. This may improve the RT simulation accuracy but still there is some degree of uncertainty associated with it. A more accurate technique to characterize the materials is reported in [13] which uses the S21 parameters obtained using the VNA and calculates a more accurate complex electrical permittivity for the specific facets defined in the RT simulation through the use of multiple-pass technique. Another way is to use crude RT calibration technique in which a set of measurements is

path. These metallic cabinets are in fact taken into consideration by the measurement process when calculating the CIR unlike in simulations where the environment description is limited to walls, doors and windows. As outcome in section IV-A, the reliability of a RT tool in LDPs modeling depends on the precision and the depth of environment description. A complete environment description in which walls, doors, windows, and furnitures are considered and properly described would lead to better simulation accuracy and avoid hence carrying laborious measurements.

1.0

Probability

0.8

0.6

0.4

0.2 Measurements RT Simulations 0.0

0

1

2

3

4

5

6

7

8

V. C ONCLUSION 9

Positioning error (m)

Fig. 3: Fingerprinting-based positioning accuracy using measurements and RT simulations respectively. collected to fine-tune the electrical parameters of the walls by using simulated annealing [14]. Notice that the amount of measurements needed for calibration/validation of RT remains significantly smaller than the amount commonly collected to train the entire FP database. B. Modeling of location-dependent parameters Modeling of LDPs is very useful in evaluating localization techniques performances. Thus, LDPs models should be as accurate as possible. Since RT tools are able to reproduce radio channel and propagation phenomena with higher accuracy, models extracted from RT simulations would be rather accurate for realistic evaluation of localization techniques. In order to assess the RT tool accuracy in LDPs modeling, we compare the RSSI and TOA values obtained using the PyRay tool and those obtained during the measurement campaign (Fig.4). In Fig. 5 we plot for AP3 and AP4 the evolution of measured and simulated RSSIs along the trajectory defined during the measurement campaign. Although the general patterns of measured and simulated RSSI look the same, some large differences occurs at some points. This is particularly explained by the fact that many pieces of furniture have not been included in the environment description. In particular, metallic cabinets (plotted in grey in Fig.4) have strong effects on radio-propagation and thus on the RSSI values. In order to reduce the gap between simulations and measurements, these pieces of furniture must be properly included in the description of the environment. As for TOA, we apply on both simulated and measured CR the Thresholded dichotomous left interval selection technique put forward in [15]. In the same way as for RSSI, we plot in Fig.6 the evolution of estimated ranging errors along the same trajectory for both simulations and measurements. These figures show that the regions with large RSSI simulation errors (in Fig.5) coincide with the regions where the ranging error obtained using measurements is very large. This is explained by the presence of metallic cabinets which obstruct the direct

In this paper we have exploited three different ray tracing tools in order to investigate the importance of using such tools in localization techniques. These investigations have pointed out two important applications: substituting/complementing laborious measurement campaigns within fingerprinting based localization techniques and modeling of location dependent parameters. Investigations and measurements-simulations comparison showed that ray tracing tools can fairly estimate these location dependent parameters once a proper description of propagation environments and antennas is performed. For both of these two applications, ray tracing tools represent a good alternative to measurement campaigns which are usually very laborious and painful to carry out. Unlike measurements, the ray tracing tools offer also the possibility of updating the description of environments and antennas according to the studied scenario. ACKNOWLEDGMENT This work was performed in the framework of the WHERE (ICT-217033) and WHERE2 (ICT-248894) projects, which are partly funded by the European Commission. R EFERENCES [1] J. McKown and J. Hamilton, R.L., “Ray tracing as a design tool for radio networks,” Network, IEEE, vol. 5, pp. 27 –30, nov. 1991. [2] S. Seidel and T. Rappaport, “Site-specific propagation prediction for wireless in-building personal communication system design,” Vehicular Technology, IEEE Transactions on, vol. 43, pp. 879 –891, nov 1994. [3] P. Meissner, D. Arnitz, T. Gigl, and K. Witrisal, “Analysis of an indoor UWB channel for multipath-aided localization,” in Ultra-Wideband (ICUWB), 2011 IEEE International Conference on, pp. 565 –569, sept. 2011. [4] M. Raspopoulos, C. Laoudias, L. Kanaris, A. Kokkinis, C. G. Panayiotou, and S. Stavrou, “3D Ray Tracing for Device-Independent Fingerprint-based Positioning in WLANs,” in 2012 Workshop on Positioning, Navigation, and Communication (WPNC), March 2012. [5] R. Burghelea, S. Avrillon, and B. Uguen, “Vector spherical harmonics antenna description for IR-UWB ray tracing simulator,” in Electromagnetics in Advanced Applications, 2009. ICEAA ’09. International Conference on, pp. 303 –306, sept. 2009. [6] B. Uguen, N. Amiot, and M. Laaraiedh, “Exploiting the Graph Description of Indoor Layout for Ray Persistency Modeling in Moving Channel,” in Proceedings of the 6th European Conference on Antennas and Propagation (EuCAP 2012), (Prague, Czech Republic), March 2012. [7] M. Varshney, Z. Ji, M. Takai, and R. Bagrodia, “Modeling environmental mobility and its effect on network protocol stack,” in Proc. IEEE Wireless Communications and Networking Conference (WCNC), (Las Vegas, NV, USA), April 2006. [8] H. Sugahara, Y. Watanabe, T. Ono, K. Okanoue, and S. Yarnazaki, “Development and experimental evaluations of ”RS-2000” - a propagation simulator for UWB systems,” in International Workshop on Ultra Wideband Systems, 2004.

35

301 297 293 111 299 295 291 AP2 109 289 107 103105 287 101 95 285 283 281 279 127 159 93 99 271273 275 277 153 155157 97 91 9 11 269 267 265 263 137 145 125 167 175 183 129 83 81 8987 85 15 13 243 241 239 237 235 233 231 189 193 197 201 207 211 229 139 147 151 161 169 177 185 17 19 37 43 49 55 61 67 73 79 115 209 205 203 199 195 191 119121 113 123 165 181 135 75 131 143 69 173 63 23 21 39 45 51 57 141 149 163 171 179 187 25 27 35 41 47 53 59 65 71 77 117 247 245 249 133 29 33 213 31 251 215 AP1 217 253 255 219 AP3 221 257 223 225 259 261 227

AP4

1

7

AP MS

Fig. 4: APs and MSs locations during the WHERE measurement campaign. −40

−40

AP3

−60 −70 −80 −90

AP4

−50

RSSI (dB)

RSSI (dB)

−50

−60 −70 −80

0

50

100

150

200

250

300

−90

0

50

Trajectory Index

100

150

200

250

300

250

300

Trajectory Index

Fig. 5: Measured versus Simulated RSSI values for AP3 and AP4. 10

10

AP3

8

Ranging error (m)

Ranging error (m)

8

6

4

2

0

AP4

6

4

2

0

50

100

150

200

250

300

Trajectory Index

0

0

50

100

150

200

Trajectory Index

Fig. 6: Measured versus Simulated TOA ranging values for AP3 and AP4.

[9] B. Uguen, M. Laaraiedh, B. Denis, J. Keignart, J. Stephan, and Y. Lostanlen, “Extraction and characterization of location-dependent UWB radio features with practical implications for indoor positioning,” in Proc. of 18th European Wireless Conference (EW2012), (Poznan, Poland), April 2012. [10] Y. Lostanlen, J. Stephan, J. Keignart, W. Wang, D. Slock, and F. Kaltenberger, “D4.1: Measurements of location-dependent channel features.” Deliverable, October 2008. [11] M. Raspopoulos, S. Stavrou, J. St´ephan, M. Laaraiedh, B. Uguen, J. Dom´ınguez, L. de Celis, B. Denis, D., and G. Agapiou, “D2.2: Intermediate: Location Information Extraction.” Deliverable, March 2012. [12] S. Stavrou and S. Saunders, “Review of constitutive parameters of building materials,” in Antennas and Propagation, 2003.(ICAP 2003). Twelfth International Conference on (Conf. Publ. No. 491), vol. 1,

pp. 211–215, IET, 2003. [13] A. Muqaibel and A. Safaai-Jazi Microwave Theory and Techniques, IEEE Transactions on, title=A new formulation for characterization of materials based on measured insertion transfer function, vol. 51, pp. 1946 – 1951, aug. 2003. [14] J. Jemai, P. Eggers, G. Pedersen, and T. Kurner, “Calibration of a UWB Sub-Band Channel Model Using Simulated Annealing,” Antennas and Propagation, IEEE Transactions on, vol. 57, no. 10, pp. 3439–3443, 2009. [15] L. Yu, M. Laaraiedh, S. Avrillon, B. Uguen, J. Keignart, and J. Stephan, “Performance Evaluation of Threshold-Based TOA Estimation Techniques Using IR-UWB Indoor Measurements,” in Proc. of 18th European Wireless Conference (EW2012), (Poznan, Poland), April 2012.

ICT–248894 WHERE2

A.6

D2.6

Model-based Evaluation of Location-based Relaying Policies in a Realistic Mobile Indoor Scenario

J. J. Nielsen, R. L. Olsen, T. K. Madsen, B. Uguen, H. P. Schwefel. Model-based Evaluation of Location-based Relaying Policies in a Realistic Mobile Indoor Scenario. In The International Workshop on Computer-Aided Modeling Analysis and Design of Communication Links and Networks (CAMAD’12), Barcelona, Spain, September 2012. c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/ republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

76 / 146

Model-based Evaluation of Location-based Relaying Policies in a Realistic Mobile Indoor Scenario Jimmy J. Nielsen⇤ , Rasmus L. Olsen⇤ , Tatiana K. Madsen⇤ , Bernard Uguen† , and Hans-Peter Schwefel⇤‡ {jjn, rlo, tatiana}@es.aau.dk, [email protected], [email protected] ⇤ Networking

and Security, Department of Electronic Systems, Aalborg University, Denmark † IETR, University of Rennes 1, Rennes, France ‡ Forschungszentrum Telekommunikation Wien - FTW, Vienna, Austria

Abstract—For WLAN systems in which relaying is used to improve throughput performance, node mobility and information collection delays can have a significant impact on the performance of a relay selection scheme. This paper analyzes this influence on the decision process using a previously developed Markov Chain model. The evaluation is done for a realistic indoor scenario that is based on ray-tracing enriched measurements from the WHERE2 project. Further, these results are compared to results obtained using an idealistic path loss model, and we show that the performance impact of node mobility and information collection delays is significantly different for the two data sets.

I. I NTRODUCTION

AP R

80 m

In WLANs it is well-known that two-hop relaying can improve throughput for some users [1], [2], [3]. However, for such link measurement based approaches as well as location based approaches [4], [5], [6] node movements and information collection delays can negatively impact performance, as shown using a Markov Chain based model in [7]. For this analysis an idealistic distance based path loss model was assumed, which did not model the multi path effects that are present in indoor environments. The present paper addresses this by comparing results obtained with such an idealistic path loss model, to results based on realistic ray tracing enhanced measurements, produced during the WHERE1 and WHERE2 projects. The results are not meant as an exhaustive site-survey but simply as a realistic comparison case. Section 2 introduces the general relay system, while the corresponding Markov model is described in Section 3. The ray-tracing enhanced measurement based use case study is described in Section 4 and in Section 5 we present a comparison of the results obtained with the idealistic and realistic data sets using the Markov Chain model from [7].

dynamic changes result purely from R’s mobility. In a locationbased relaying approach, the AP in this situation needs to take a decision whether – based on its inaccurate (delayed) knowledge of the position of the candidate relay – relaying is beneficial or not. The mapping of the estimated position of R to the relay decision is called a relay policy, here represented by ⇡(ˆ xR , yˆR ) 2 {R, D}, where R is ’relay’ and D is direct transmission. In this paper, we assume that such relay decision is taken just before each individual data fragment transmission. This scenario could correspond to a use case where a user is located in an office with quite poor wi-fi coverage, where he tries to exploit bypassing colleagues’ wireless devices as mobile relays. It should be noted that the proposed modeling approach works equally well for the situation where the relay is static and instead the destination is mobile.

Queue delay MAC + transmission delay

D

AP

Relay 

...



Measurement unicast Measurement transmission scheduled 80 m 50 m

Fig. 1. System with static AP, mobile relay (R), and static destination (D). The gradient ellipse illustrates where relayed transmissions yield higher throughput than direct transmissions. The dashed circle around R is the position uncertainty.

Fig. 2. Location measurements are transmitted periodically with rate ⌧ and are subject to queueing delay as well as MAC and transmission delay. The latter two are represented by a distribution with mean 1/µ.

II. S YSTEM D ESCRIPTION We consider the three node system sketched in Fig. 1, consisting of a static access point (AP), mobile relay R, and static destination D. The candidate relay R has a position (xR (t), yR (t)) that changes over time. Note that D is static, so This work has been performed in the framework of the ICT project ICT248894 WHERE2, which is partly funded by the European Union. The Telecommunications Research Center Vienna (ftw.) is supported by the Austrian Government and by the City of Vienna within the competence center program COMET.

The positions of the static AP and of the destination D are assumed to be known at the AP, hence only the mobile relay node will periodically (with rate ⌧ ) send position updates (ˆ xR , yˆR ) to the AP. It is assumed that such location information is available at R through, e.g., a GPS system. In order to investigate the impact of the forwarding delays of positioning information in mobile scenarios, it is assumed that the only cause of inaccurate information is the mobility. Therefore, the position information obtained at R is assumed to be exact.

was used in [7]: Transitions are only allowed to the neighboring grid states and all states have the same overall state leaving rate µm . As a consequence, the average movement speed of the candidate relay can be readily obtained as v¯ = d/µm , where d is the distance between neighboring grid-points. The example in Fig. 3(a) (b) shows an example mobility model for indoor scenarios in which a wall blocks certain movements. The actual used more complicated setting will be later explained in Figure 5(a). m

m 1

2

m m

m

m

3

m

m N+2

m

m m

m

m

m

m

2N+1

2N+2

m

m m

N+3

2N

m

m

m

m

m

m

(N-1) *N+1

N

m

m

m

N+1

m

m

Wall

The location measurements are transmitted from the relay to the AP, as sketched in Fig. 2: Node R obtains an exact coordinate of its current location, wraps it into a WLAN packet and passes it to its WLAN interface. At the WLAN interface, there could be a queuing delay until the location message reaches the first position in the (finite) interface queue, followed by a subsequent MAC and transmission delay. The sum of MAC and transmission delays are assumed to show a distribution with mean 1/µ. The AP’s estimate of Node R’s position is based on the last received location measurements. Since the relay is mobile its true position may differ from the AP’s estimate, depending on the stochastic mobility model of the candidate relay. For many mobility models, the older the most recent measurement becomes, the less accurate the AP’s view is expected to be. Depending on the AP’s belief on the relay’s location it will choose to either make a relayed (R) or direct data transmission (D). The resulting throughput will depend on this choice. A throughput increase can be achieved by the choice of a high bit rate coding scheme when the relay to node distances are shorter than the AP to destination distance. However, as we assume decode and forward relaying here, the relayed transmission requires two packets to be sent which will lead to additional channel occupation affecting throughput. For the numerical results later, we utilize the throughput model of [8], which is summarized in Section IV-C.

(N-1) *N+2

m

m m

2N+3

3N

(N-1) *N+3

N2

(a) Mobility model

(b) Mobility model with wall

AP view = D

 2 D/-



·wR

·wD



·wR

·ploss

1 -/-

·ploss

III. M ODEL FOR R ELAYING P ERFORMANCE A SSESSMENT For optimal performance, it is desirable to make the choice that maximizes the overall achieved throughput. In the considered scenario, this optimal choice depends on the node mobility models, on the distant-dependent propagation characteristics, and on the strategy (period of the location updates) and forwarding delays (queuing, MAC and transmission) of these location updates. In this section, we summarize the Markov model we developed in [7] that is later in this paper applied to a realistic indoor scenario. The main approach for the Markov model is to start from a product-space representation of two parts: 1) a continuous time Markov model for the spatial mobility of the candidate relay node (the ’true’ coordinates); 2) a model of location update procedures and of the resulting AP view. As these two parts are not completely independent, a pure product space approach however is not sufficient, but requires subsequent modifications as summarized in this section. Performance metrics are calculated from the steady-state solution of the Markov chain, see [7] for more details.

4 D/D

·wD

3 R/-

·wD

5 D/R

6 R/D ·(1-ploss)

·wR

·(1-ploss)

·ploss ·(1-ploss)

AP view = R

·(1-ploss) ·ploss

·(1-ploss)

·ploss

9 D/-

·wD



·wR

11 D/D

·(1-ploss)

·wR ·ploss

8 -/-



·wD

7 R/R

·wD

12 D/R

13 R/D

10 R/·wR 

14 R/R

Geographical grid point 1 Geographical grid point 2 ...

(c) Information forwarding model

Fig. 3. (a) Example Markov mobility model without obstacles or walls. N ⇥ N discrete grid-points represent the geographic space. µm is the stateleaving rate. (b) A similar mobility model, but with a wall. (c) State overview for Markov model of information forwarding.

Markov Mobility Model First element of the relaying Markov model is a continuous time Markov model that describes the candidate relay’s stochastic mobility. The geographic 2-dimensional space is discretized, for instance via a equidistant grid. The states then represent the current true position of the candidate relay within the grid. Transition rates between the states characterize the mobility. Fig. 3(a) shows a base model without obstacles as it

AP View and Information Forwarding In order to model the AP view on the relaying input information and the location update process, the state-space at each coordinate (state of the mobility model) an extended model of each grid point was developed in [7] as shown in Fig. 3(c). The following gives a very brief summary of the model, see [7] for more details. Based on the AP’s

current knowledge of the mobile relay’s position the AP can apply the relaying policy that maps geographic coordinates to either action, ’D’=Direct Transmission or ’R’=Relay. Instead of keeping the coordinate knowledge in the state-space, the Markov model just tracks the current decision resulting from the policy, i.e. in the example direct transmission in States 1-7 or relay transmissions in States 8-14. Updates on the relay’s position are created by the relay node with rate ⌧ , shifting the state to the right in Figure 3(c), indicating one (middle column) or two (right column of states) updates in progress. Depending on whether the policy maps the update in progress to a direct or a relay view, the change will go up (for direct, e.g. 2 to 4) and down (for relay, e.g. 3 to 7). The up or down is determined by the functions wD and wR which attains either the value 0 or 1 depending on the grid position and the policy ⇡(m). The model in the figure allows for two updates in progress, representing a interface card buffer with room for one update (and one being in progress of transmitted). Subject to the network delay rate µ the update message reaches the AP, which may lead to a change of decision for the AP. Finally, with probability ploss a message may be dropped by the network. Geographic Throughput Model In order to calculate expected throughput of the relaying system, we here assume that this expected throughput is only influenced by variability due to mobility. As we here assume that AP and target nodes are static, we only require the candidate node’s position as input. The throughput for a the direct transmission is thereby constant, while the relayed cased depends on the location of the relaying node. The specific WLAN 802.11 throughput model from [8] is used in the case study; it will be summarized in Section IV-C. IV. M EASUREMENT BASED U SE C ASE S TUDY For demonstrating the application of the proposed model, we consider a case study that reflects the scenario in Fig. 1. AP and D Nodes are static, whereas the R Node moves according to the Markov mobility model presented in Section III. Note that the AP and D positions have been manually selected in order to show a scenario where relaying can be beneficial. For other positions of the AP and D nodes the direct transmission mode is always best, which would result in a valid but very trivial analysis outcome. The Nodes are equipped with 802.11g based radios, but modified to support relaying as mentioned in [4]. Furthermore, since the area is small compared to the typical range of 802.11g the transmission parameters have been scaled down to imitate a scenario in which relaying is usable. Table I lists the used scenario and simulation parameters. The specific scenario is described in the following. A. Ray Tracing Simulation PyRay tool, an UWB ray tracing simulator has been used to produces a large set of UWB received signal on a grid of 51 pseudo Access-points ⇥ 363 Mobile Station (MSs) on a uniform grid which cover the same office building where the

UWB WHERE1 measurement campaign has been conducted. Those simulated received waveforms have been built in order to reproduce as closely as possible the whole transmission chain including various effects (waveform convolution, Tx and Rx antennas) in order to allow a fair comparison with actual measurement data. For simplicity, this paper only covers the leftmost 2/5 of the building shown in Fig. 4. Now, given the UWB received signals, in order to evaluate the scenario under the assumption of IEEE 802.11g WiFi, it is necessary to first extract the narrow band path loss from the measurements, and thereafter estimate the achievable throughput from the path loss.

Fig. 4. The indoor Layout with furniture and the used grid of point (pseudo AP and destination nodes are in red.)

B. Evaluating Narrow Band Path Loss from UWB Signal The UWB waveform p(t) which has been used for simulation is energy normalized and is expressed mathematically as follow: s p 2 t 2 2 p e Tp cos(2⇡fc t) p(t) = (1) Tp ⇡ with Tp =

2 B

dB

r

dB

ln(10) , 20

(2)

where fc = 4.49 GHz, B 3dB = 500 MHz , dB = 3. Thus, what is available, in the database, for each link is the computed convolution product r(t) = p(t) ? h(t) of the UWB waveform p(t) and the CIR of the channel impulse response h(t). In practice this convolution is performed in the frequency domain using the Fourier transform of each quantity R(f ) = P (f )H(f ). In order to obtain a location dependent parameter, which is independent of the shape and magnitude of the applied waveform, it is necessary to define a path loss quantity constructed from the UWB received waveform. The used path loss is defined as the energy ratio between the applied and the received signal filtered out in a narrow band of interest which is evaluated as follows: 0R 1 fc + 2b 2 b |P (f )| df B fc 2 C LNB (fc , b) = 10 log10 @ R (3) A fc + 2b 2 b |R(f )| df fc 2

For our simulation the value b = 20 MHz has been chosen, corresponding to B = N ⇥ b with N = 25, and fc = 2.412 GHz corresponding to IEEE 802.11g channel 1. For an extensive description of the ray tracing scenario, see [9].

C. Used Throughput Model

A. Candidate Policies

Given the extracted path loss for a narrow band corresponding to IEEE 802.11g channel 1, we estimate the achievable throughput from the path loss. The throughput model of IEEE 802.11 that is used in this paper is based on previous work, further detailed in [8]. As the throughput is given by Delivered data Transmission time , and using bit-error-rate models to calculate the frame error probabilities (see [8] for details), the throughput can be calculated as: Psuc · BMSDU Sdir = (4) E[Ttx ]

For evaluation we consider two location based policies, one which requires an accurate location estimate (grid accuracy) and another which relies on a coarser room-level accuracy. For comparison we consider also the two static policies of always transmitting directly or always using the relay. Since the information collection has an impact on the performance when using the location based policies, we will for the accurate location based policy consider both a delayed information collection and the ideal case with instantaneous collection. The following are the combinations we consider:

where Psuc is the probability of a successful MAC layer frame delivery, E[Ttx ] is the duration of a MAC frame delivery attempt, and BMSDU is the MAC payload size given in octets. In the following, we use the indices 1 and 2 to indicate the AP-r and r-d transmissions. The throughput for the two-hop relaying algorithm is calculated as: Srel =

pri,1 pri,2 sec,1 sec,2 Psuc + Psuc Psuc ) · BMSDU (Psuc

E[Ttxpri,1 ] + E[Ttxpri,2 ] + E[Ttxsec,1 ] + E[Ttxsec,2 ]

(5)

.

Name Ideal Locally optimal Heuristic Always direct Always relay

Loc. accuracy grid level grid level room level -

Info. collection instantaneous delayed delayed -

For the heuristic scheme we have defined that it will only use the relay within the two rooms within the rectangle between the two corners (8;1) to (11;6).

The throughput model is used to estimate the transmission throughput functions TD (m) and TR (m) for each of the M grid points, indexed by m. 12

0

11

Path loss measurements PL = −51.6 + 10*2.0219*log10(d/d0)

10

−10

9 8

(a) Measurement based throughput (b) Path loss model based throughput model model

Path loss [dB]

−20

7 6 5

−30

Fig. 6. AP and D positions are marked by a black square (12, 4) and black cross (5, 1), respectively. The color of each grid point shows the achievable relay throughput, if the mobile relay is in that position. The green circles mark the grid points in which the relay throughput exceeds the direct throughput.

−40 4 3

−50

2

−60 0 10

1 0

0

1

2

3

4

5

6

7

8

9

10

11

(a) Mobility model

12

13

14

1

10 distance d [m]

(b) Path loss model estimation

Fig. 5. (a) Mobility model in room layout. Red lines are possible movements between states (blue dots) and black dashed lines are walls. (b) Estimation of path loss model parameters, namely reference path loss at 1m and path loss exponent.

V. M ODEL BASED A NALYSIS In addition to the ray tracing based data set, we consider a comparison case where a simple path loss model is used to estimate the path loss at different positions in the building. We have used a path loss model of the form [10]: P L = P Ld0 + 10n log10 (d/d0 ) [dB]

(6)

Any multi path effects caused by the building structure are thus not accounted for. Initially, we fitted this model to the distribution of all measured path losses in the dataset relative to the AP point. Specifically, we used a Least Summed Squared Error optimization to obtain the parameters P Ld0 = 51.6 dB and n = 2.022, shown in Fig. 5(b). Furthermore, the transmit power was adjusted to 4.32 mW so that the direct transmission yielded the same throughput as with the measurement data.

Parameter Measurement delay rate µ Network loss probability ploss Noise floor Ricean K Data frame payload BMSDU Average movement speed v ¯ Measurement update rate ⌧

Value 105 s 1 0 85 dBm 6 1500 bytes 0.5 m/s 1s 1

TABLE I D EFAULT SCENARIO PARAMETERS .

B. Results and Discussion For evaluating the performance of the different schemes described in the section above, we have applied the mobility model shown in Fig. 5(a), the two throughput models shown in Fig. 6 and the default parameters listed in Tab. I on the Markov Chain model described in Section III. The results obtained when varying the relay movement speed, location information update interval and network delay are presented in the Fig. 7, 8, and 9. Common for all result plots is that the Ideal, Always direct, and Always relay policies are constant, since they are not depending on the varied parameters.

7 6.5

6

Avg. throughput [Mbit/s]

Avg. throughput [Mbit/s]

7 6.5

5.5 5 4.5 Ideal Locally optimal Heuristic Always direct Always relay

4 3.5 3

1

2 3 mobility speed [m/s]

4

6 5.5 5 4.5

3.5 3

5

2 3 mobility speed [m/s]

4

5

Average throughput for different policies for varying relay speed.

7

7 6.5

6

Avg. throughput [Mbit/s]

Avg. throughput [Mbit/s]

1

(b) Path loss data

6.5

5.5 5 4.5 Ideal Locally optimal Heuristic Always direct Always relay

4 3.5 3

Ideal Locally optimal Heuristic Always direct Always relay

4

(a) Ray-Tracing data Fig. 7.

VI. C ONCLUSIONS AND O UTLOOK

2

4 6 update interval τ−1 [s]

8

6 5.5 5 4.5 Ideal Locally optimal Heuristic Always direct Always relay

4 3.5 3

10

(a) Ray-Tracing data

2

4 6 update interval τ−1 [s]

8

10

(b) Path loss data

Fig. 8. Average throughput for different policies for varying measurement update rate.

7

7

6.5

6.5

6

Avg. throughput [Mbit/s]

Avg. throughput [Mbit/s]

For all parameters being varied with the locally optimal scheme, the ray tracing data results show a bigger impact on performance. This is due to the large variations in achieved throughput between neighbor grid points, caused by multipath effects in the indoor environment. The path loss data set, on the other hand, does not show the same rapid change between neighbor grid points and the impact of mobility, low update rate and network delays on the locally optimal scheme is therefore less pronounced. Actually, the impact seems identical for the two schemes with the path loss data set, whereas with the ray-tracing data set the heuristic algorithm is hardly impacted, but is also constantly below the Always direct scheme. The heuristic room based scheme does not seem to be a good choice in indoor multi path environments. Looking specifically at the mobility speed results for the locally optimal scheme in Fig. 7, the ray-tracing data set shows that it is possible to achieve higher throughputs compared to the path loss data set, for mobility speeds in the range of walking speeds (0.5 - 1.5 m/s).

5.5 5 4.5 Ideal Locally optimal Heuristic Always direct Always relay

4 3.5 3

1

2 3 Network delay µ−1 [s]

4

(a) Ray-Tracing data

6 5.5 5 4.5 Ideal Locally optimal Heuristic Always direct Always relay

4 3.5

5

3

1

2 3 Network delay µ−1 [s]

4

5

(b) Path loss data

Fig. 9. Average throughput for different policies for varying network delay.

Given extracted narrow band path loss information from a database of ray-tracing enhanced measurements of a realistic indoor scenario, this paper presents a Markov Chain (MC) based performance evaluation of location based relaying policies, assuming IEEE 802.11g wireless network equipment. The MC model, proposed in [7] takes into account the inaccuracies caused by a mobile relay’s possible movements as well as information collection and network delays. Our results compare the achieved performance in the ray tracing based data set with the performance achieved using a simple distance based path loss model. We find that while the overall impact of mobility and delays is well accounted for, the variation in performance between neighboring grid points in the indoor building map for the ray tracing data set, due to multi path propagation effects, leads to a significant difference in the results obtained with the two input data sets. A possible improvement could be to artificially include the statistical variation in the path loss model data set. Due to the measurement/ray tracing data set being used is geographically quite limited in size, it has been necessary to scale down the parameters of the wireless communication, in order to properly showcase a scenario where relaying can be beneficial. As scaling the parameters might introduce some differences compared to considering a larger area, this would be an obvious task for future work. R EFERENCES [1] J.-S. Liu and Y.-C. Wong, “A relay-based multirate protocol in infrastructure wireless lans,” in Networks, 2004. (ICON 2004). Proceedings. 12th IEEE International Conference on, vol. 1, 2004, pp. 201 – 206 vol.1. [2] H. Zhu and G. Cao, “rDCF: A relay-enabled medium access control protocol for wireless ad hoc networks,” IEEE Transactions on Mobile Computing, pp. 1201–1214, 2006. [3] P. Liu, Z. Tao, S. Narayanan, T. Korakis, and S. Panwar, “CoopMAC: A cooperative MAC for wireless LANs,” IEEE Journal on Selected Areas in Communications, vol. 25, no. 2, p. 340, 2007. [4] J. J. Nielsen, T. K. Madsen, and H.-P. Schwefel, “Location-based mobile relay selection and impact of inaccurate path loss model parameters,” in Proc. of the IEEE Wireless Communications and Networking Conference - WCNC, 2010. [5] B. Zhao and M. Valenti, “Practical relay networks: a generalization of hybrid-arq,” IEEE Journal on Selected Areas in Communications, vol. 23, no. 1, pp. 7–18, 2005. [6] M. Zorzi and R. Rao, “Geographic random forwarding (GeRaF) for ad hoc and sensor networks: multihop performance,” Mobile Computing, IEEE Transactions on, vol. 2, no. 4, pp. 337–348, 2004. [7] J. J. Nielsen, R. L. Olsen, T. K. Madsen, and H.-P. Schwefel, “On the impact of information delay on location-based relaying: A markov modeling approach,” in Proc. of the IEEE Wireless Communications and Networking Conference - WCNC, 2012. [8] J. J. Nielsen, T. K. Madsen, and H.-P. Schwefel, “Location-based relay selection and power adaptation enabling simultaneous transmissions,” in Proc. of the GLOBECOM Workshops, 2010. [9] M. Laaraiedh, B. Uguen, J. Stephan, Y. Corre, Y. Lostanlen, M. Rapopoulos, and S. Stavrou, “Ray tracing based radio simulator modeling for indoor localization purposes,” Proceedings of Computer Aided Modeling and Design (CAMAD), 2012. [10] G. Durgin, T. Rappaport, and H. Xu, “Measurements and models for radio path loss and penetration loss inand around homes and trees at 5.85 GHz,” IEEE Transactions on Communications, vol. 46, no. 11, pp. 1484–1496, 1998.

ICT–248894 WHERE2

A.7

D2.6

Common Scenario Simulation Storage

Marios Raspopoulos.

82 / 146

Common Scenario Simulation Storage Various simulations were carried out in the defined common scenario using various narrowband and UWB Simulators and the impulse responses for each receiver location with respect to each Transmitter were computed. All the results have been exported into MATLAB and a separate MATLAB structure was created for each Transmitter. Each datafile is saved as a separate file having the following filenames:  1st Narrow Band Simulation using Sigint’s RT Simulator (3DTruNET) Narrow_RT_Sigint_TXN  2nd Narrow Band Simulation using Sigint’s RT Simulator (3DTruNET) Second_Narrow_RT_Sigint_TXN  Simulation using SIRADEL’s simulator SIRADEL_TxN ,where N indicates the number of the respective transmitter. The structure of the results is shown in Figure 1. It contains various details about the simulation parameters and the IR received at the receiver location an omnidirectional antenna but also the 8 IRs received by directional antennas rotated by 45 degrees each time.

Figure 1: MATLAB structure for every Transmitter

To allow easy dissemination of the simulation results an MSSQL schema has been defined, has been integrated to the WHERE2 database and has been populated with the results. In this context, Figure 2 shows the schema of the part of the database that is used to hold simulated data in the synthetic environment which have been generated using Ray Tracing Simulations by SIGINT Solutions, SIRADEL and University of Rennes.

Transmitters PK

TxID

INTEGER

FK1

simulationID TxName TxX TxY TxZ TxPower TxFrequency TxBW TxType

INTEGER CHAR(50) DOUBLE DOUBLE DOUBLE CHAR(10) DOUBLE DOUBLE CHAR(50)

Simulation omniResults PK

omniResultsID INTEGER

FK1 FK2

RxID TxID RxPower omniIR

INTEGER INTEGER DOUBLE CHAR(2000)

PK

sectorResults

simulationID

INTEGER

PK

sectorResultsID INTEGER

simulationDescription simulatorName noOfReflections noOfTransmissions noOfDiffractions simulationEnvironment

CHAR(2000) CHAR(50) INTEGER INTEGER INTEGER CHAR(100)

FK1 FK2

RxID TxID lowerAngle upperAngle IR sectorPower

INTEGER INTEGER DOUBLE DOUBLE CHAR(2000) DOUBLE

Receivers PK

RxID

INTEGER

FK3

simulationID RxName RxSensitivity RxType RxX RxY RxZ numberOfSectors

INTEGER CHAR(50) DOUBLE CHAR(50) DOUBLE DOUBLE DOUBLE INTEGER

Figure 2: Database schema for Task 2.3 Ray Tracing simulations

A comprehensive description of the tables defined in this schema is provided below: Simulation Table: It is the parent of the schema and contains general information about the simulation like description, simulator name, simulation environment and number of interactions considered in the simulation (reflections, transmissions and diffractions)

Entry Name

Data Type

simulationID

Integer

simulationDescription

Char(2000)

Description Uniquely identifies the simulation. Private Key Description of the simulation

simulatorName

Char(50)

The name of the simulator used to generate the database

noOfReflections

Integer

No of reflections considered in the simulator (-1 means unlimited)

noOfTransmissions

Integer

No of transmissions considered in the simulator (-1 means unlimited)

noOfDiffractions

Integer

No of diffractions considered in the simulator (-1 means unlimited)

simulationEnvironment

Char(100)

Description of the simulation Environment

Transmitters Table Contains all the transmitters defined in the simulation. The contained information includes the transmitter name, coordinates, transmitted power, bandwidth, frequency and antenna type. Entry Name

Data Type

Description

TxID

Integer

Uniquely identifies the Transmitter. Private Key

simulationID

Integer

Simulation to which the Transmitter is linked to. Foreign Key

TxName

Char(50)

The Name of the Transmitter

TxX

Double

The X coordinate of the transmitter (in meters)

TxY

Double

The Y coordinate of the transmitter (in meters)

TxZ

Double

The Z coordinate of the transmitter (in meters)

Char(10)

The transmit power of the Transmitter (in dBm)

TxPower TxFrequency

Double

The frequency of transmission in Hz

TxBW

Double

The Bandwidth of transmission in Hz

TxType

Char(50)

The Type of the Transmitter antenna (e.g. isotropic, dipole ...)

Receivers Table Contains all the Receivers defined in the simulation. The contained information includes the receiver name, coordinates, receiver sensitivity and antenna type.

Entry Name

Data Type

Description

RxID

Integer

Uniquely identifies the Receiver. Private Key

simulationID

Integer

Simulation to which the Receiver is linked to. Foreign Key

RxName RxX

Char(50) Double

The Name of the Receiver The X coordinate of the Receiver (in meters)

RxY

Double

The Y coordinate of the Receiver (in meters)

RxZ

Double

The Z coordinate of the Receiver (in meters)

RxSensitivity

Double

The sensitivity of the Receiver (in dBm)

RxType

Char(50)

The Type of the Receiver antenna (e.g. isotropic, dipole etc.)

omniResults Table Contains the received parameters (Rxpower and impulse response) assuming that the receiver is omnidirectional. Each row represents the results of a specific RX-TX pair which are provided as foreign keys in the table. The RxPower field is a DOUBLE field which defines the received power in dBm. The omniIR is defined as a string for simplicity. The format is defined in the table below. Entry Name

Data Type

Description

omniResultsID

Integer

Uniquely identifies the results received by the Receiver assuming that the receiver is omnidirectional. Private Key

RxID

Integer

Receiver ID in which the sector belongs to. Foreign Key.

TxID

Integer

Transmitter ID from which the signal was received. Foreign Key.

RxPower

Double

The received power receiver by the Receiver from the transmitter identified by TxID (in dBm) The received omnidirectional Impulse Response received by the receiver from transmitter TxID. It is defined as a string (for hard disc space issues and simplicity) with the following format (comma separated values): “CIR, NoOfComponents, time1, componentPower1, time2, componentPower2,…, timeN, componentPowerN” 

omniIR

Char(2000)

  

CIR  identifies that this is a Channel Impulse Response NoOfComponents  is an integer which identifies how many received paths constitute the impulse response TimeX  time (in nanosec) of the received component (path) componentPower  the power (in dBm) of the received component (path)

e.g. for 3 components: “CIR,3,0, -78,10,-59,20,-60”

sectorResults Table Similar to the omniResults table but each entry represents the received signal parameters for a specific sector of the receiver antenna.

Entry Name

Data Type

Description

sectorResultsID

Integer

Uniquely identifies the results received by a specific sector of the receiver of the Receiver. Private Key

RxID

Integer

Receiver ID in which the sector belongs to. Foreign Key.

TxID

Integer

Transmitter ID from which the signal was received. Foreign Key.

lowerAngle

Double

The lower angle (in degrees) of the receiver sector

upperAngle

Double

The upper angle (in degrees) of the receiver sector

Power

Double

The received signal power (in dBm) by the sector. The received Impulse Response received by the receiver (RxID) sector from transmitter TxID. It is defined as a string (for hard disc space issues and simplicity) with the following format (comma separated values): “CIR, NoOfComponents, time1, componentPower1, time2, componentPower2,…, timeN, componentPowerN”

IR

Char(2000 )

   

CIR  identifies that this is a Channel Impulse Response NoOfComponents  is an integer which identifies how many received paths constitute the impulse response TimeX  time (in nanosec) of the received component (path) componentPower  the power (in dBm) of the received component (path)

e.g. for 3 components: “CIR,3,0, -78,10,-59,20,-60”

In order to populate this part of the database with the Ray Tracing Simulations carried out in Task 2.3 by SIGINT Solutions and SIRADEL the following steps need to be followed: 

Attached the database as described in Deliverable D4.4



Download the whole folder from here https://bscw.eurecom.fr/bscw/bscw.cgi/387801



You need to run “RunThisToLoadToSQL.m” but first you need to change the first line in the file which defines the connection string to the SQL database (depends where you have the database installed and the credentials you have defined for the user). For example: connectionString='PROVIDER=SQLOLEDB; Data Source=MARIOS2PC\SQLEXPRESS; initial catalog=NEWWHERE2; User ID=where2; password=where';

Basically it is needed to change the Data Source, the User ID and the password.

ICT–248894 WHERE2

A.8

D2.6

Rectangular Room Dimensions Estimation Using Narrowband Signal and Sectorized Antennas

Igor Arambasic, Francisco Javier Casajus Quiros, Ivana Raos. Proceedings of the 5th International Symposium on Communications Control and Signal Processing (ISCCSP), 2012, pp.1-6, 2-4 May 2012 c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

88 / 146

Proceedings of the 5th International Symposium on Communications, Control and Signal Processing, ISCCSP 2012, Rome, Italy, 2-4 May 2012

RECTANGULAR ROOM DIMENSIONS ESTIMATION USING NARROWBAND SIGNAL AND SECTORIZED ANTENNAS Igor Arambasic, Francisco Javier Casajus Quiros, Ivana Raos Departamento de Señales, Sistemas y Radiocomunicaciones E.T.S.I. de Telecomunicación, Universidad Politécnica de Madrid, Madrid, Spain E-mails: [igor, javier, ivana]@gaps.ssr.upm.es ABSTRACT A system for estimation of unknown rectangular room dimensions based on two radio transceivers, both capable of full duplex operations, is presented. The approach is based on CIR measurements taken at the same place where the signal is transmitted (generated), commonly known as selfto-self CIR. Another novelty is the receiver antenna design which consists of eight sectorized antennas with 45° aperture in the horizontal plane, whose total coverage corresponds to the isotropic one. The dimensions of a rectangular room are reconstructed directly from radio impulse responses by extracting the information regarding features like round trip time, received signal strength and reverberation time. Using radar approach the estimation of walls and corners positions are derived. Additionally, the analysis of the absorption coefficient of the test environment is conducted and a typical coefficient for office room with furniture is proposed. Its accuracy is confirmed through the results of volume estimation. Tests using measured data were performed, and the simulation results confirm the feasibility of the approach. Index Terms —absorption coefficient, CIR, narrowband, reverberation time, room volume, RTT, sectorized antenna, self-to-self CIR, TOA 1. INTRODUCTION Most of the existing methods for estimating the room volume are based on Channel Impulse Response (CIR) between two points located inside the room. This type of observation is in acoustics commonly known as room impulse response (RIR). Actually, different methods for extraction of room dimensions out of RIR have been proposed over the years, being all sensitive to transmitreceive positions and room reflection coefficients [1-4]. None of these approaches combines RIR with self-to-self CIR which provides more information obtained at the same place, and as such should increase the stability of the solution. Further on, several studies reported in the recent literature attempted to address this problem with radio signals based on Ultra-Wide-Band (UWB) techniques. Some adopt a battype UWB radar approach [5], other require a highly

978-1-4673-0276-0/12/$31.00 ©2012 IEEE

specific hardware configuration, for example central transmitter with receive antennas array [6] or monostatic CIR measurements (i.e. self-to-self channel sounding) [7]. Regardless of different techniques, bottom line is that all of them require large bandwidths and as such introduce significant hardware restrictions. Contrary to UWB methods, we propose to shift hardware constraints from bandwidth to the design of the receiver antennas. Hence, the bandwidth is reduced to narrow-band (100MHz) and isotropic antenna is replaced with eight sectorized antennas with 45° aperture in the horizontal. Eventually, using self-to-self CIR obtained at each one of these eight antennas the presence of walls and corners can be evaluated utilizing round trip time (RTT) measurements. The rest of the paper is organized as follows. The details of synthetic test environment used for simulations are given in section 2. Afterwards, system model is described. The process of estimating the positions of the walls is presented in section 4 while two last sections are reserved for simulation results and conclusions. 2. SYNTHETIC EVALUATION ENVIRONMENT During the first year of the ICT-WHERE2 project [8], partners in charge of developing indoor mapping, selflocalization and mobility learning algorithms expressed the need for the common synthetic evaluation framework in order to ensure fair performance assessment and comparison of the proposed algorithms. This common synthetic environment is based on measurements collected during the ICT-WHERE1 [9] project which are complemented by deterministic raytracing simulations of CIRs calculated by three partners of WHERE2 for three different bandwidths (UWB, Wide Band and Narrow Band). It represents a typical indoor office environment, as specified in section 6.1.5 of deliverable D1.1b [10]. The size of the simulated environment is approx 30x12m, and transmitters and receivers height was set to 1.5m. It should be mentioned that CIRs where calculated for two types of receiver (Rx) antennas. The transmitting (Tx) antenna is always isotropic but the receiver can be set to isotropic, or consists of eight sectorized Rx antennas with 45° aperture in the horizontal plane. Eventually, these sectors correspond to dividing the received signal into eight

sections and covering 360º azimuth angles. However, this also means that for each pair of Tx-Rx links nine CIRs are calculated (one per sector antenna and one per omnidirectional). Eventually, Rx positions are defined with a grid of approximately 1mx1m density (all together 363 positions as seen in Figure 1) while Tx positions are located on the same grid but their number is reduced to 51 to ease the computational load.

2. This hypothesis is true if we consider four sectors of Rx antenna to be orthogonal to the walls meaning their TOA calculations correspond to radar reading. In other words, half of self-to-self TOA calculation of the corresponding sector antenna matches the distance of the transceiver to the wall orthogonal to that section.

Figure 2: Different measurement corrections are required depending on observed distance for corner calculation Figure 1: Rx positions inside common synthetic environment The simulations resulted in CIRs for all Tx-Rx links and all three bandwidths. Out of these CIRs, different location/context-radio parameters, such as Received Signal Strength Indicators (RSSIs), multipath Times of Arrival TOAs, (Average) Power delay Profiles ((A)PDPs), channel delay spread, or mean excess delay, etc., can be extracted. In this paper synthetic narrow-band environment obtained by SIGINT 3DTruEM software [11] was used, and RTT or TOA of the strongest peak, RSS parameters and reverberation time were exploited. The simulations have been carried out at 3.5GHz (this is the frequency that the Ray Tracing Simulator was calibrated for) and a bandwidth of 100MHz considering an unlimited number of refractions and up to one diffraction, as defined in [12]. 3. SYSTEM MODEL One important requirement for this model is the ability of both transceivers to support self-to-self CIR, meaning that they are able to transmit and receive simultaneously. Utilizing self-to-self CIR on sectorized Rx antenna enables a variety of applications like self-localization, room shape detection, room volume determination and eventually the solution to simultaneous localization and mapping (SLAM) problem which is our final goal. In this preliminary work we restrict the research to determining the dimension of the rectangular room and its volume. When representing rectangular room as geometrical model walls are seen as reflecting lines in a twodimensional space, while corners are characterized by double reflections at two orthogonal walls as seen in Figure

One the other hand, four remaining sectors of Rx antenna are pointing to the corners and since this includes double reflection their reading of distance should be systematically corrected. The correction should be smaller for the corners far away from the transceiver and larger as the transceiver is located closer to the corner as pointed out with a circle in Figure 2. This correction parameter is labelled as C and is set empirically to values between 0.7 and 1 depending on the calculated distance from the wall. C=0.7 matches distance measurements between 0m and 1.5m, C=0.75 measurements between 1.5 and 3m, C=0.8 measurements between 3 and 4.5m etc. with maximum C equally 1. Apart from correction parameter C, two other prerequisites are necessary for the approach; one is accurate path loss model for calculating the distance between the transceivers and the other is faithful room absorption factor for determining the room volume. Both parameters are determined empirically, by fitting them to the true values obtained over the entire synthetic environment scenario except for one room which is left as “unknown” for testing the method. Eventually, seven rectangular rooms (marked with circles in Figure 3.) were used for extracting these parameters and then the approach is tested on the room marked with the diamond. For path loss model, one slope approximation model for the relationship between attenuation and distance has been used:

L( f c ) = L0 ( f c ) + 10n log10

d d0

(1)

Where d0 is the unit distance (1m) and L0(fc) its corresponding attenuation at transmission frequency fc.

Figure 3: Training rooms marked with ellipse and test room marked with diamond. It was tested for 7 rectangular rooms, taking into account all possible pairs of Tx-Rx links inside the corresponding room. Finally L0=-35.85 and n=1.7 produce the best fitted approximation to be used for this path loss model inside synthetic narrow-band environment. These values are consistent with the ones found in [13] and [14] for LOS short-range indoor n value (1.87 and 1.5, respectively). In order to stress the importance of the third factor, the absorption coefficient, it should be noted that the shape of the self-to-self CIR is determined by: room volume determine the tail of CIR absorption (or reflection) coefficient of the walls and materials inside the room determine the slope of CIR transceiver position inside the room determine the position of maximum CIR value A combination of volume, surface and absorption coefficient is often found in room acoustic in form of another parameter known as reverberation time ( RT60 ) :

RT60 =

( )

4 ln 10 6 V c aS

(2)

where c is speed of light, V is room volume, S is room surface, and a is absorption coefficient. RT60 is actually defined as a decay time that takes a signal to drop 60dB below its original level [15,16]. In our environment RT60 can be calculated out of isotropic self-toself CIR or out of isotropic CIR between the two transceivers located inside the same room. In both cases RT60 is calculated the same way as time difference between the maximum of CIR and 60dB decay from the maximum, as shown in Figure 4. Notice that the tail of CIR presents unstable measurements (dots); so in order to accurately determine the 60dB decay; it is approximated by polynomial function.

RT60 using polynomial

Figure 4: Calculation of

approximation on the CIR’s tail This is done by finding the coefficients of a polynomial p(x) of first degree (N=1) that fits the data in a least squares sense.

p( x ) =

N n =0

p N − n +1 x n

(3)

Larger degree for the approximation is not needed as the tail power is smooth function. Afterwards, it is interpolated by a factor of 10 to obtain improved time precision (see line). Hence, if we calculate RT time, for all available self-to-self CIRs inside the seven rectangular rooms, by knowing the volume of the corresponding room, and its surface, according to equation (2) we could obtain its actual absorption factor.

a = 1.842 × 10 − 7

V RT60 S

(4)

The histogram of all absorption coefficients is presented in Figure 5. It shows that even though rooms are of different sizes, include different furniture, and not all have walls built out of same materials, the absorption factor is concentrated between 0.4 and 0.6. Eventually, the average value (a=0.492) of the obtained absorption coefficients is proposed when calculating the volume of an unknown midsize room with furniture.

corner, right upper corner and middle of the bottom wall). Its furniture composed of metallic bookcase (on the left wall), one large wooden desk and one table. The room is not perfectly rectangular with approximate size of 5.1mx7.3m with 2.4m height. The data obtained for CIR of the transceiver 325 at position 334 is seen in Figure 7. It consists of nine CIRs. The central one corresponds to isotropic antenna, while eight surrounding CIRs match the corresponding sectorized antenna of the Rx.

Figure 5: Estimated pdf of absorption coefficient 4. ALGORITHM DESCRIPTION The focus of the approach is determining the dimensions of the rectangular room. To do so, at least two transceivers are necessary. Legend:

325

transceiver 334

7.32m

table bookcase

5.13m

Figure 6: layout of test room with bookcase, tables and two transceivers The position of the transceivers inside the room is unknown, but their orientation is predefined as 4 sectorized Rx antennas are orthogonal to the walls. Actually, the influence of antenna orientation with respect to walls should still be investigated, but if found critical the solution would be to replace Rx sectorized antennas with 2D array of eight isotropic antennas. As a result beam-forming would be used for determining the correct orthogonal orientation with respect to the walls (minimum RTT measurements). When extracting different location/context-radio parameters two standard CIR obtained between the transceivers, and two self-to-self CIR, are available. The algorithm will be presented by applying the approach to two transceivers located at the positions marked with filled circle (see Figure 6). These two circles correspond to positions inside the grid tagged as 325 and 334. The test room is actually a reception hall which explains the existence of three doors (left upper

Figure 7: CIRs between transceivers 325 (Tx) - 334 (Rx) The algorithm starts by calculating the distance between the transceivers, out of isotropic standard CIRs, using path loss model. The model was calibrated in another environment using the same transceivers, as already explained in section 3. Afterwards, sector CIRs of Rx antennas of both transceivers are used for detecting the direction of the emitter. This is easily seen in Figure 7 where upper-left sector receives highest power level when compared to other sectors. Eventually this statement constructs two wall observations with respect to the observed sector. In case of a corners calculations (sectors not orthogonal to the walls), the obtained distance is multiplied by correction coefficient C, and then two corner observations are produced. The same calculation is done on the other transceiver. The resulting angle is the average of the two readings if their opposite sectors are at least neighbouring sectors. Otherwise, the measurements should be discarded as useless. This case was not encountered inside narrow-band synthetic environment but should not be discarded in real life measurements. The precision of angle classification is of 22.5º. After these two steps the transceivers are positioned with respect to each-another enabling the data obtained out of self-to-self CIRs to be constructive. The third step consists of locating the round trip time of the strongest peak for self-to-self CIR of the sectorized RX antenna, at both positions. Usually the first peak is also the strongest one, but

as the room is not empty the first peak is not selected since its reflection source might be the furniture and not the wall.

part with 12 observations for each wall is seen in Figure 8. Four correspond to the directly to orthogonal wall observations, and eight correspond to the two corners. These 12 observations are then projected on the straight line as seen in Figure 9 and median of all projections produces the wall position estimate. Once the locations of four walls are established, only the height of the room is missing in order to be able to determine the volume. If we calculate the reverberation time for two self-to-self isotropic CIRs, and two standard isotropic CIRs, four (V/S) relations can be obtained out of (2). Then, by solving equation (5), four height values are produced, and the average value is delivered as height estimate.

V XˆYˆZ = S 2 XˆYˆ + XˆZ + YˆZ

(

)

(5)

where Xˆ and Yˆ are estimated floor dimensions, V is volume, S is surface and Z is height.

Figure 8: Estimation of walls and corners positions based on RTT of self-to-self CIRs obtained at sectorized Rx antenna

Figure 10: CDF of relative estimation error of room dimensions X and Y 5. SIMULATION RESULTS

Figure 9: Left wall estimation example In case of narrow-band environment (100MHz), TOA values are multiples of 10ns which matches round-trip distance to the wall to be a multiple of 1.5m. Hence, it could be said that, if RTT of the orthogonal section is 10ns, the wall is positioned between 1.5m and 3m from the transceiver. Eventually these readings (1.5m and 3m) would be the two points the corresponding Rx sectorized-antenna produces. Hence, when determining the wall position we

The CDF of relative error obtained for dimension of X and Y are depicted in Figure 10. The results correspond to room 1, of Siradel building, with size of 5.1mx7.3m. This office room includes metallic bookcase and 2 wooden tables. It is not perfectly rectangular, and consists of one glass wall and three wooden doors. As such the environment can be considered as rather hostile. Nevertheless, the CDF show that in more than 80% of calculations the relative error of the obtained dimensions is bellow 15%, while in 50% of calculations the error would be less than 11%. After obtaining two room dimensions, these can be used together with the average absorption factor for determining

room height and consequently room volume. The performance of the proposed approach can be seen in Figure 11. The CDF shows that in more than 80% of calculations the relative volume error is bellow 20%.

8. REFERENCES [1] M. Kuster, “Reliability of estimating the room volume from a single room impulse response,” J. Acoust. Soc. Am., vol. 124, no. 2, pp. 982–993, Aug. 2008. [2] Shabtai, N.R.; Zigel, Y.; Rafaely, B.; "Feature selection for room volume identification from room impulse response," WASPAA '09. IEEE Workshop on Applications of Signal Processing to Audio and Acoustics, 2009., pp.249-252, Oct. 2009 [3] Shabtai, N.R.; Zigela, Y.; Rafaely, B.; , "Estimating the room volume from room impulse response via hypothesis verification approach," IEEE/SP 15th Workshop on Statistical Signal Processing, 2009. SSP '09, pp.717-720, Sept 2009 [4] J. B. Allen and D. A. Berkley, “Image method for efficiently simulating small room acoustics,” J. Acoust. Soc. Am., vol. 65, no. 4, pp. 943–950, Apr. 1979. [5] Seitz, J.; Schaub, M.; Hirsch, O.; Zetik, R.; Deissler, T.; Thoma, R.; Thielecke, J.; , "UWB feature localization for imaging," IEEE International Conference on Ultra-Wideband, 2008. ICUWB 2008, vol.2, pp.199-202, 10-12 Sept. 2008

Figure 11: CDF of relative volume estimation error This is in line with relative errors of X and Y dimensions, and means that no additional error was introduced in this step. Eventually, it confirms that the average absorption coefficient is a good choice for this kind of room. 6. CONCLUSION AND FUTURE WORK The algorithm for indoor localization with 100MHz RF signal is presented. The main distinctive property of the algorithm is the spatial discrimination of the received signal and absorption factor usage. It is shown that the even with 100MHz signal (up to 3m localization error per estimation due to sampling rate) the relative position and volume estimation error can be maintained in more than 80% of points below 15% and 20% respectively. This confirms that the algorithm is appropriate and can be used for indoor localization and volume estimation. For distance estimation the simple one-slope path loss model was found to be adequate, with attenuation and path loss exponent parameters estimated empirically from the data. Further work includes analysis of improvements with respect to precision and sensibility using larger bandwidth signal (up to 500MHz) and/or introducing more transceivers. Additional research on localization and tracking should also be valuable for SLAM problem solution. 7. ACKNOWLEDGMENTS This work has been carried out in the frame of the WHERE2 (FP7-ICT 248894) project, which is funded by European Union, and Spanish national MCIN project TEC200914219-C03-01.

[6] Deissler, T.; Thielecke, J.; , "Feature based indoor mapping using a bat-type UWB radar," 2009. IEEE International Conference on Ultra-Wideband, 2009. ICUWB, pp.475-479, Sept. 2009

[7] Wenyu Guo; Filer, N.P.; , "On the Accuracy of an Indoor Location-sensing Technique Suitable for Impulse Radio Networks," IEEE International Conference on Communications, 2007. ICC '07, pp.3987-3992, 24-28 June 2007 [8] http://www.ict-where2.eu/. [9] “Measurements of location-dependent channel features," technical report, Deliverable D4.1 of WHERE1 Project (ICT217033), Oct. 2008. [10] “Scenarios and parameters”, Deliverable D1.1b of the WHERE2 Project (ICT-248894), August 2011. [11] “Ray-tracing tools for dynamic positioning," technical report, Deliverable D1.5 of WHERE2 Project (ICT-248894), July 2011. [12] “Self-learning positioning using inferred context information (intermediate report)," tech. rep., Deliverable D2.3 of WHERE2 Project (ICT-248894), Dec. 2011. [13] “D1.1.2 WINNER II channel models–part I: channel models,” WINNER Information Society Technologies, Tech. Rep. IST-4027756 WINNER II, version 1.2, Sep. 2007. [14] Bultitude, R.J.C.; Hahn, R.F.; Davies, R.J.; "Propagation considerations for the design of an indoor broad-band communications system at EHF," IEEE Transactions on Vehicular Technology, vol.47, no.1, pp.235-245, Feb 1998 [15] H. Kuttruff, Room Acoustics, Fourth Edition, London, Spon Press, 2000 [16] Andersen, J.B.; Nielsen, J.O.; Pedersen, G.F.; Bauch, G.; Herdin, M.; , "Room electromagnetics," Antennas and Propagation Magazine, IEEE , vol.49, no.2, pp.27-33, April 2007

ICT–248894 WHERE2

A.9

D2.6

Self-Positioning and mapping of rectangular rooms with sectorized narrowband antennas

Igor Arambasic, Marios Raspopoulos, Javier Casajus Quiros, Ivana Raos, Stavros Stavrou 20th International Conference on Software, Telecommunications and Computer Networks (SoftCOM), 2012 , Split, Croatia, September 2012. c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

95 / 146

Self-Positioning and mapping of rectangular rooms with sectorized narrowband antennas Igor Arambasic1, Marios Raspopoulos2, Javier Casajus Quiros1, Ivana Raos1, Stavros Stavrou2 1

Departamento de Señales, Sistemas y Radiocomunicaciones E.T.S.I. de Telecomunicación, Universidad Politécnica de Madrid, Madrid, Spain E-mails: [igor, javier, ivana]@gaps.ssr.upm.es 2 Sigint Solutions Ltd, Nicosia, Cyprus E-mail: [m.raspopoulos, s.stavrou]@sigintsolutions.com

Abstract: A system for simultaneous 2D estimation of rectangular room and transceiver localization is proposed. The system is based on two radio transceivers, both capable of full duplex operations (simultaneous transmission and reception). This property enables measurements of channel impulse response (CIR) at the same place the signal is transmitted (generated), commonly known as self-to-self CIR. Another novelty of the proposed system is the spatial CIR discrimination that is possible with the receiver antenna design which consists of eight sectorized antennas with 45° aperture in the horizontal plane and total coverage equal to the isotropic one. The dimensions of a rectangular room are reconstructed directly from spatial radio impulse responses by extracting the information regarding round trip time (RTT). Using radar approach estimation of walls and corners positions is derived. Tests using measured data were performed, and the simulation results confirm the feasibility of the approach.

1. INTRODUCTION In this paper, we propose a novel system, which addresses positioning and mapping problems of rectangular rooms utilizing inferred context information from measurements of wireless communication properties. The idea of simultaneously obtaining location and environment map has received considerable attention in robotics in the last 30 years inside Simultaneous Localization and Mapping (SLAM) problem. It was originally developed by Hugh DurrantWhyte and John J. Leonard [1] based on earlier work by Smith et al. [2]. The procedure consists of collecting the measurements and incremental building of the map, within an unknown environment without any a priori knowledge. Since it is more a concept type than a specific algorithm, different SLAM techniques have been used and proposed in the literature. Their extensive overview can be found in [3, 4]. We believe that these techniques from robotics can be extended to the context of wireless anchor-less positioning where anchor-less refers to the ability of the devices to infer their own coordinates, in a local relative system, without any fixed reference node, base station or access point. Several studies reported in the recent literature attempted to address SLAM problem with radio signals based on Impulse Radio Ultra-Wide-Band (UWB) techniques [5-9] or

a bat-type UWB radar approach [10]. Other contributions requiring a highly specific hardware configuration have also been proposed, for example central transmitter with receiver antennas array [11] or monostatic CIR measurements (i.e. self-to-self channel sounding) [12]. Most of the approaches rely on the geometrical interpretation of the arrival times of the resolved echoes. These echoes are assumed to result from simple electromagnetic interactions (e.g. simple-bounce reflections on the walls), enabling a unique correspondence between the pattern of path arrivals and nodes positions. Simple relationships are invoked to link elements of geometry (including relative nodes locations) with the observed multipath inter-delays and to deliver a four-wall map of the room. Regardless of approaches differences, bottom line is that all of them require large bandwidths and as such introduce significant hardware restrictions. Contrary to UWB methods, we propose to shift hardware constraints from bandwidth to the design of the receiver antennas. This is achieved by limiting the system to narrowband (NB) signals (100MHz) and introducing eight sectorized antennas with 45° horizontal aperture. By using self-localization, the signal emitted at omnidirectional is received at eight sectorized antennas. The presence of walls and corners is evaluated using RTT measurements. Some preliminary results of this approach are published in [13], focused on determining all three dimensions of an unknown rectangular room. In this contribution 2D mapping algorithm is improved by detecting the position of transceivers close to corners and by carrying out self-localization on two cooperating transceivers. This is seen as simplification of SLAM problem where transceiver is a part of an “intelligent” mobile robot platform that is updating its positions and the room map based on the observations obtained at consecutive locations. Here, the number of consecutive locations is limited to two. The performance of the approach is evaluated under realistic synthetic test environment based on NB ray tracing simulations. 2. DESCRIPTION OF THE ENVIRONMENT Synthetic test environment of ICT-WHERE2 project [14] is based on measurements collected during the ICTWHERE1 [15] which are complemented by detailed 3D

description of the site geometry and deterministic ray-tracing simulations of CIR calculated for three different bandwidths (UWB, Wide Band (WB) and NB). It represents a typical indoor office environment, as specified in section 6.1.5 in deliverable D1.1b. [16]. The size of the simulated environment is approximately 30x12m. Transmitters (Tx) and receivers (Rx) heights are set to 1.5m. Rx positions are defined in a grid of approximately 1mx1m density, as shown in Figure 1, with a total of 363 positions. Tx positions are located on the same grid but their number is reduced to 51 to ease the computational load. Tx antenna is always isotropic while CIRs where calculated for two types of Rx antennas: isotropic and eight sectorized 45° horizontal aperture antennas.

orientation inside an unknown environment should still be investigated, but if found critical the possible solution could be to replace Rx sectorized antennas with 2D array of eight isotropic antennas. As a result, beam-forming could be used for detecting antenna sections perpendicular to the walls (minimum RTT measurements). Hence, the orientation of the transceivers is predefined with four sectors of Rx antenna orthogonal to the walls. This means that their TOA calculations correspond to typical radar reading where half of self-to-self TOA calculation matches the distance of the transceiver to the wall orthogonal to that section. Sectors of Rx antennas that are pointing to the corners may receive double reflection signals. Therefore, their reading of distance should be systematically corrected as explained in [13]. The correction parameter is labeled Cp and is set empirically to values between 0.75 and 1 depending on the calculated distance from the wall:

⎧ d ⎪⎪ 0.75 + , d ≤ 7.5m 30 Cp = ⎨ ⎪ 1 , d > 7.5m ⎪⎩

(1)

where d is distance obtained from the RTT as: Figure 1 - Rx positions inside common synthetic environment In this paper synthetic narrow-band environment obtained by SIGINT 3DTruEM software was used [17]. The software is calibrated at 3.5GHz central frequency and 100MHz bandwidth using real measurements data. Unlimited number of refractions and up to one diffraction were taken into account. The emitting power of transmitter was set to 10dBm, while the receiver sensitivity was set to -110dB, as defined in [18]. 3. ALGORITHM IMPLEMENTATION The algorithm is developed to solve self-localization problem of two cooperating nodes located inside unknown rectangular room. As a result of localization the distances of the nodes with respect to room walls are provided allowing the construction of room 2D map. Transceivers support full duplex self-to-self CIR, meaning they are able to transmit and receive simultaneously. Signal is emitted by omnidirectional antenna while reception is available with eight sectorized antennas. The mutual distance and orientation of the nodes is assumed to be known. This is an initial assumption in order to obtain the performance bounds of the algorithm. When representing rectangular room with geometrical model, walls are seen as reflecting lines in a two-dimensional space, while corners are characterized by double reflections from two orthogonal walls. The influence of initial antenna

d=

RTT ⋅ c 2

(2)

with c being the speed of light. The relation in (1) is obtained by fitting the corner distance estimation to true corner distance of 35 Tx positions located in 8 rectangular rooms of the synthetic environment. Each of the received reflections is actually sum of different reflection signals due to indoor multipath propagation. Because of the the receiver time resolution, multiple delayed signal versions cannot be discriminated and they are modeled with a single path (single reflection). This simplification impacts RTT approximation, but its impact is limited through the diversity, that is, introduction of second transceiver that also performs the room dimensions estimations. The RTT is defined here as the time required for the strongest peak to reach the transceiver. Usually this is also the first peak inside CIR but, since the room is not empty, the first peak is not selected, as its reflection source might be the furniture and not the wall. Since the bandwidth is limited to 100MHz the precision of CIR’s time scale is 10ns, which corresponds to resolution of 1.5m. The test room (marked with ellipse in Figure 1) is a reception hall with three doors (upper left corner, upper right corner and middle of the bottom wall). Its furniture is composed of metallic bookcase (on the left wall) and two wooden tables. The room is not perfectly rectangular with approximate size of 5.1mx7.3m and 2.4m height. The algorithm is applied to two transceivers, one located in the upper left corner and other close to room center.

(corresponds to cos(45°) or sin(45°)). Afterwards, the obtained RTT distance is modified by coefficient Cp before producing the wall observations:

d5..8 =

c 2 RTTjci C p (d) 2 2

(5)

c 1 2 d9..12 = (RTTjci C p (d) + ) 2 BW 2 with superscript ci ,i=1..2, corresponding to two corner sections of each of the j=1..2 transceivers.

Figure 2 - self-to-self CIRs of transceiver at upper position Eight CIRs data, obtained at transceiver located close to upper left corner are seen in Figure 2. Each CIR corresponds to one Rx sector antenna, which is pointing in the direction of the position of sub-image. In other words, the first line of 3 images would correspond to upper left corner, upper wall and upper right corner CIRs. When determining the wall position we part from three observations for each wall obtained at both transceivers (one direct radar reading and two corner estimations). This produces six RTT calculations per wall. However, since the precision of RTT readings is directly dependent on bandwidth this must be taken into account:

d=

c RTT, 2

c⎛ 1 ⎞ ⎜ RTT + ⎟ 2⎝ BW ⎠

(3)

where BW is signal bandwidth. In other words if RTT of the orthogonal section, obtained with 100MHz bandwidth, is 20ns, we cannot claim that the wall is exactly at 3m distance, but we could presume it is positioned between 3m and 4.5m from the transceiver. Eventually this statement constructs two distance values for each RTT observation. By incorporating this strategy, the overall number of wall observations is increased to twelve. Four correspond to radar readings:

d1..2 =

c RTTjo 2

c⎛ 1 ⎞ d3..4 = ⎜ RTTjo + ⎟ 2⎝ BW ⎠

(4)

where superscript o stands for orthogonal section, and subscript j=1..2 correspond to transceivers count. In case of corner calculations, if we initially assume that the antenna pointing to the corner is actually pointing at 45° angle with respect to two perpendicular walls, a simple trigonometric calculation produces the wall distance by multiplying corner reading by scaling factor

2 2

Figure 3: Estimation of walls and corners positions based on RTT of self-to-self CIRs obtained at sectorized Rx antenna These 12 estimations (d1..12) are graphically described in Figure 3 where two transceivers are marked with filled circles. Their relative position with respect to each other is known (employing sensors implemented at robot platform) and is depicted with dashed line. The arrows represent possible, estimated positions of walls and corners. The length of arrows corresponds to the precision of RTT readings (c/2BW). Before producing the final wall estimation two constraints, that are used for discrimination of inadequate estimation, are put on each observation: 1) If the transceiver is in a corner its distance with respect to each of corner walls cannot be smaller than 0.25m and larger than 1.5m. We assume that the transceiver is in a corner if two orthogonal readings, together with the corner observation between them, produce the maximum energy reading at 0ns. 2) If both transceivers are inside the room, walls cannot be in-between. If any of these constraints is not met, the distance estimation is regarded as unreliable and is not taken into account. The

mean value of the rest of observations produces the wall estimation. 4. SIMULATION RESULTS The results for two transceivers located inside office room, marked with ellipse in Figure 1, are presented in Figure 4. The upper transceiver is set at position (0,0) and all other measurements are done with respect to this point. As can be seen, the estimated room dimensions are slightly larger than the real ones. This could have been expected as we opted not to base RTT calculation only on the first CIR peak, but to use the maximum peak and afterwards to add bandwidth precision to produce the second estimation with the same importance weight. Lower vertical error presented in this example, can be a result of line-of-sight conditions between the transceiver and vertical walls at both transceivers location. In case of horizontal walls, the bookcase at left path, and table on right path are seen as the source of higher horizontal errors of the transceiver located in the center of the room.

real estimated transcievers furniture

Figure 5 - CDF of relative horizontal and vertical error [%] When observing the localization error of the transceivers inside the room, we focus on absolute error since relative error of corner estimations (which are close to the wall) can produce misleading conclusions. The absolute error of transceiver distance to each wall is presented in Figure 6. The results show the effectiveness of presented NB estimation procedure of two cooperating transceivers, as the localization error is kept below 1.2m in 95% of cases. It can be observed that the maximum error is below 1.8m, but the errors over 1.2m can be regarded as outliers (occurs in 10% of scenarios). It is also interesting to notice the different CDF error slope that is more abrupt for smaller errors, as in 75% of cases the error is below 0.7m. This large amount of values with low errors support the claim that this approach is adequate for positioning inside the rectangular rooms with furniture.

Figure 4 - Example of wall estimation and Tx positioning The cumulative density function (CDF) of relative error obtained for horizontal and vertical 2D room dimension are depicted in Figure 5. The simulations were done for more than 80 different pairs of transceiver positions inside the same office room. The CDF show that in over 85% of calculations the relative error of the obtained dimensions is bellow 15%, while in 50% of calculations the relative error of room dimensions would be less than 7%. When compared to the results presented in [13] the two constraints implemented in this algorithm shift the error slope to the left by approximately 5%. The CDF confirm that the estimation of horizontal walls is susceptible to higher errors as furniture covers higher percentage of these walls.

Figure 6 - CDF of absolute positioning error

5. CONCLUSION The algorithm for indoor localization and 2D mapping of rectangular rooms with spatial discrimination and 100MHz RF signal is presented. The algorithm is carrying out selflocalization with two cooperating transceivers. The spatial discrimination of the received signal is possible with sectorized multiple receiver antennas whose total coverage equals the isotropic antenna coverage. The method is tested using synthetic data obtained with deterministic ray-tracing simulations of CIRs. The test room is not perfectly rectangular, and consists of one glass wall, two brick ones, two wooden doors and furniture. It is shown that the even with 100MHz signal (up to 1.5m localization error per estimation due to sampling rate) the relative mapping error can be maintained in more than 85% of points below 15%, while in 50% of calculations the relative error of room dimensions would be less than 7%. By detecting the position of transceivers close to corners and by excluding the presence of walls between the transceivers, the improvement of 5% of relative estimation error is obtained when compared to the results in [13]. The absolute positioning error with respect do real wall distance is below 0.6m in more than 70% of calculations. This confirms that the algorithm is appropriate and can be used for indoor localization and mapping of rectangular rooms with furniture. Further work would include applying realistic mobility model and SLAM analysis. Additional research to estimate the Rx antenna sector orientation should also be valuable for SLAM problem solution. 6. ACKNOWLEDGMENTS This work has been carried out in the frame of the WHERE2 (FP7-ICT 248894) project, which is partly funded by the European Union and in the frame of Spanish MCIN project TEC2009-14219-C03-01. REFERENCES [1] J.J. Leonard and H.F. Durrant-Whyte. “Mobile robot localization by tracking geometric beacons”. IEEE Transactions on Robotics and Automation, 7(3):376 –382, jun 1991. [2] R. Smith, M. Self, and P. Cheeseman. “Estimating uncertain spatial relationships in robotics”. In. Proceedings of IEEE International Conference on Robotics and Automation. 1987, volume 4, page 850, mar 1987. [3] H. Durrant-Whyte and T. Bailey. “Simultaneous localization and mapping: part i.” IEEE Robotics & Automation Magazine, 13(2):99–110, 2006. [4] T. Bailey and H. Durrant-Whyte. “Simultaneous localization and mapping (slam): Part ii”. Robotics & Automation Magazine, IEEE, 13(3):108–117, 2006.

[5] Barton S. K. Guo W., Filer N. P. “2d indoor mapping and location-sensing using an impulse radio network”. In IEEE ICU’05, Zurich, pages 296–361, Sept 2005. [6] Filer N. P. Barton S. K. Guo W., Thomson S. L. Knowledge base assisted mapping for an impulse radio indoor location-sensing technique. In International Workshop on Wireless Ad-hoc Networks 2005, London, may 2005. [7] Barton S. K. Guo W., Filer N. P. A 2d uwb indoor wireless technique for location-aware applications. In 1st International Symposium on Broadband Communications (ISBC’04), Harrogate, page 58, Dec 2004. [8] Barton S. K. Guo W., Filer N. P. “A novel wireless mapping and positioning technique for impulse radio networks”. In 18th triennial URSI International Symposium on Electromagnetic Theory, Pisa 2004, volume 2, pages 712– 714, may 2004. [9] Filer N. P. Guo W. 2.5d indoor mapping and location sensing using an impulse radio network. In IEE Seminar on Ultra Wideband Systems, Technologies and Applications 2006, London, pages 211–215, April 2006. [10] Seitz, J.; Schaub, M.; Hirsch, O.; Zetik, R.; Deissler, T.; Thoma, R.; Thielecke, J.; , "UWB feature localization for imaging," IEEE International Conference on UltraWideband, 2008. ICUWB 2008, vol.2, pp.199-202, 10-12 Sept. 2008 [11] Deissler, T.; Thielecke, J.; , "Feature based indoor mapping using a bat-type UWB radar," 2009. IEEE International Conference on Ultra-Wideband, 2009. ICUWB, pp.475-479, 9-11 Sept. 2009 [12] Wenyu Guo; Filer, N.P.; , "On the Accuracy of an Indoor Location-sensing Technique Suitable for Impulse Radio Networks," IEEE International Conference on Communications, 2007. ICC '07, pp.3987-3992, 24-28 June 2007 [13] I. Arambasic, J. Casajus Quiros and I. Raos. “Rectangular Room Dimensions Estimation Using Narrowband Signal and Sectorized Antennas”, 5th International Symposium on Communications, Control and Signal Processing, Rome, Italy, May 2012. [14] http://www.ict-where2.eu/. [15] “Measurements of location-dependent channel features," technical report, Deliverable D4.1 of the WHERE1 Project (ICT-217033), Oct. 2008. [16] “Scenarios and parameters”, Deliverable D1.1b of the WHERE2 Project (ICT-248894), August 2011. [17] “Ray-tracing tools for dynamic positioning," tech. rep., Deliverable D1.5 of the WHERE2 Project (ICT-248894), July 2011. [18] “Self-learning positioning using inferred context information (intermediate report)," tech. rep., Deliverable D2.3 of the WHERE2 Project (ICT-248894), Dec. 2011.

ICT–248894 WHERE2

A.10

D2.6

Anchor-less Self-Positioning in Rectangular Room Based on Sectorized Narrowband Antennas

Igor Arambasic, Javier Casajus Quiros, Ivana Raos, Marios Raspopoulos, Stavros Stavrou Proceedings of 19th European Wireless Conference (EW2013), Guildford, UK, April 2013 c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

101 / 146

Anchor-less Self-Positioning in Rectangular Room Based on Sectorized Narrowband Antennas Igor Arambasic, Javier Casajus Quiros, Ivana Raos

Marios Raspopoulos, Stavros Stavrou

E.T.S.I. de Telecomunicación Universidad Politécnica de Madrid Madrid, Spain [igor, javier, ivana]@gaps.ssr.upm.es

Sigint Solutions Ltd, Nicosia, Cyprus [m.raspopoulos, s.stavrou]@sigintsolutions.com

Abstract— A system for anchor-less self-positioning and simultaneous 2D estimation of rectangular room dimensions is proposed. The system is based on narrowband transceiver capable of simultaneous transmission and reception. The transmission is generated at omnidirectional antenna, while eight sectorized antennas with 45° apertures in horizontal are used for reception. This full duplex property enables measurements of spatially discriminated channel impulse response (CIR) at the same place where the signal is originated. Hence, by using radar approach, the presence of walls and corners can be evaluated based on round trip time (RTT) measurements. The novel contribution of this paper includes position estimation confidence indicator defined as instantaneous received power of the CIR’s strongest element. Additionally, we assume that a transceiver is a part of an “intelligent” mobile robot platform capable of moving, updating its positions, and keeping track of past positions together with the corresponding measurement data. Here, we propose a new mobility model, focused on efficient exploration of an unknown environment by moving towards the walls with lower sum of corresponding confidence indicators. Simulations inside narrowband synthetic test environment are performed and the performance of localization and room dimension estimates, as a function of number of observations, is evaluated. Index Terms—full duplex antennas, channel impulse response, indoor, mapping, mobility model, radar approach estimation, ray tracing, rectangular rooms, round trip time, simultaneous localization and mapping, sectorized narrowband antennas, selfpositioning, spatial CIR discrimination, transceiver localization

I. INTRODUCTION The problem of Simultaneous Localization and (environment) Mapping (SLAM) has received considerable attention in robotics in the last 30 years. It includes a device capable of moving and collecting some kind of measurements at different positions. These measurements are used to incrementally build the map of the environment without any a priori knowledge. SLAM concept was originally developed in [1, 2]. Over the years, different techniques have been used and proposed in literature. Their extensive overview can be found in [3, 4]. We believe that these techniques from robotics can be extended to the context of wireless anchor-less positioning where anchor-less refers to the ability of the devices to infer their own coordinates, in a local relative system, without any fixed reference node, base station or access point. The solution

presented in this paper is based on inferred context information obtained from measurements of wireless link properties. Several studies reported in the recent literature attempted to address this problem with radio signals based on Ultra-WideBand (UWB) techniques [5-12]. These approaches, similar to ours, rely on the geometrical interpretation of the arrival times of the resolvable echoes. These echoes are assumed to result from simple electromagnetic interactions (e.g. simple-bounce reflections on the walls), enabling a unique correspondence between the pattern of path arrivals and nodes positions. Contrary to UWB methods, in [13,14] we proposed to shift hardware constraints from bandwidth to the design of the receiver antennas. By doing so, the bandwidth is reduced to narrow-band (100MHz) and isotropic receiver antenna is replaced with eight sectorized antennas with 45° aperture in the horizontal. Eventually, the presence of walls and corners is evaluated utilizing round trip time (RTT) measurements of signal emitted at omnidirectional and received at eight sectorized antennas. In contrast to [13, 14], in this contribution the transceiver is seen as a part of an “intelligent” mobile robot platform that is updating its position and the room map based on the observations obtained at consecutive locations. Self-positioning and 2D mapping algorithm is improved when compared to [14] by introducing the measure of confidence index in calculations and by proposing a mobility model designed to efficiently explore the unknown scenario. The performance of the approach is evaluated under realistic synthetic test environment based on NB ray tracing simulations. The rest of the paper is organized as follows. Selfpositioning algorithm and mobility model are described in sections 2 and 3 respectively. Synthetic test environment used for simulations is explained in section 4, while two last sections are reserved for simulation results and conclusions. II. SELF POSITIONING ALGORITHM Anchor-less self-positioning approach is based on the narrowband transceiver (100MHz), which consists of one omnidirectional and 8 sectorized antennas. Additionally the transceiver should support full duplex self-to-self CIR, meaning it is able to transmit (omnidirectional antenna) and receive (sectorized antenna) simultaneously. By implementing

self-to-self CIR on sectorized antenna a variety of applications are enabled. Our aim here is to add the solution to SLAM problem to this growing list that already includes room volume determination [13] and self-localization of two cooperating nodes [14]. It should be emphasized at this stage that the use of sectorized antennas is not essential for this method as this kind of antennas can be seen as 2D array of N omni-directional antennas capable of forming a beam with 45° aperture in horizontal plane. Naturally, the algorithm is not restricted only to 45° aperture and if this aperture is narrower the positioning results should be more accurate. Beam-forming approach could also be used for defining the initial orientation of the transceiver, that is, for detecting the pairs of beams perpendicular to the walls. This can be achieved by scanning the 360° in horizontal and selecting the pairs of beams with minimum RTT measurements. We define RTT as time required for the strongest peak to reach the transceiver. The first peak was not selected since the room is not empty and the source of its reflection might be the furniture and not the wall.

single path since they cannot be discriminated due to receiver time resolution. Nevertheless this ambiguity is taken into account by making RTT reading dependent on transceiver’s bandwidth, and its impact is limited through spatial diversity (repetition of CIR measurements at different positions inside the room). As described in [14], each observation distance (d) in fact consist of two estimations, one corresponding to simple RTT readings and the other to distance ambiguity of the estimation:

d=

c f (RTT ), 2

c ( f (RTT ) + f (BW )) 2

(1)

where c is speed of light, BW is signal bandwidth. With signal bandwidth restricted to 100MHz the precision of CIR’s time scale is 10ns, which corresponds to distance resolution of 1.5m (0.5c/BW). Hence, as shown in eq. (1), for the 20ns RTT measurement, we do not claim that the wall is exactly at 3m distance, but we presume it is positioned between 3m and 4.5m from the transceiver. Eventually every node position offers six estimation of each wall, four corresponding to corner estimations and two to direct wall positions. These six estimations for one wall can be mathematically described as:

c RTTno 2 c! 1 $ m d 2n = # RTTno + & 2" BW % m d 1n =

2 c RTTnc1 C p (d) 2 2 c 1 ) d m4n = (RTTnc1 C p (d) + 2 BW 2 c m d 5n = RTTnc2 C p (d) 2 2 c 1 m = (RTTnc2 C p (d) + ) d 6n 2 BW m = d 3n

Fig. 1. Example of eight sectorized self-to-self CIRs

When performing self-positioning we part from eight CIR observations, obtained at initial arbitrary position inside the room. It is assumed that the selection of adequate antenna orientation is performed by physically rotating the mobile platform and selecting the position with minimum RTT readings. Hence four sectorized antennas are always perpendicular to the four walls and four are pointing to the corners. This is seen in Fig. 1. where each CIR corresponds to one sector antenna pointing in the direction of the position of sub-image (i.e. the first column of 3 sub-images would correspond to upper left corner, left wall and down left corner CIR). Self-localization is achieved by calculating node distances with respect to four room walls. The walls are modeled as reflecting lines in a two-dimensional space, while corners are characterized by double reflections from two orthogonal walls. However, in real life, each received reflections is actually a sum of different reflection signals as a result of indoormultipath propagation. These reflections are modeled with a

2 2

(2)

2 2

where n stands for n-th observation, m is the room wall (m=1..4). RTTno corresponds to sectorized antenna perpendicular to the wall, and RTTnc1 , RTTnc2 to the measurements of two sectorized antennas pointing to two corners. Cp is the corner distance correction parameter introduced in [13] and calculated in dependency on simple corner distance measurement as: " d $$ 0.75 + , d ≤ 7.5m 30 Cp = # (3) $ 1 , d > 7.5m $% where d is distance obtained from the RTT as: RTT ⋅ c (4) d= 2 The scaling factor 2 2 , which appears in corner calculations, comes from the hypothesis that the antenna

pointing to the corner is actually pointing at 45° angle with respect to two perpendicular walls. Hence, a simple trigonometric operation produces the wall distance by multiplying corner reading by this scaling factor (corresponding to cos(45°) or sin(45°)). If relative positions of the transceivers at the moment CIRs are obtained are known, wall estimation data can be successfully combined to produce adequate position estimates. The geometrical description of estimations in equation (2) is explained in detail in [14] where mean value of all estimates was used for producing the final result. The novel part of selflocalization algorithm presented in this contribution is introduction of distance estimation confidence indicator (W), which is defined as the instantaneous received power of the CIR’s strongest element (corresponding to RTT measurement):

W = 10

max(CIR) 10

(4)

We also define the sum of all confidence indicators for one wall, as an indicator of wall estimation precision as W0

W0m =

∑ Wi,nm

(5)

i=1:6 n=1:N

m i,n

where W describes m-th wall and i-th estimation of n-th observations of the total number of N observations (i=1..6, m=1..4, n=1..N). By taking into account the measure confidence indicator, the final distance estimate of wall m is seen as: Wm m d m = ∑ i,nm di,n (6) i=1:6 W 0 n=1:N

The positioning algorithm is done in a way to provide four wall estimates with respect to the actual transceivers position by taking into account past positions, and their corresponding wall estimates. Since the result of self-localization problem is provided in form of distances with respect to room walls, the construction of room 2D map is straightforward and is obtained simultaneously with localization parameter as in any other SLAM approach. III. MOBILITY MODEL Apart from controlling the movement and orientation, the “intelligent” robotic platform is also responsible for selecting movement trajectory in order to maximize the positioning precision. We part from the fact that the position information is more precise if all walls are located more precisely. On the other hand, if just one wall is located poorly, the estimated transceiver’s position will most likely carry a significant error. Hence, it makes sense to move towards this wall and explore its position with more accuracy. As transceivers true position, nor the actual positions of the walls are known, we suggest to use the sum of distance confidence indicators of each wall ( W0m ) for selecting the one with lowest precision estimate. The presented mobility model is based on the environment built over a rectangular grid of equally spaced positions. If

none of possible movement is obstructed (due to wall or furniture) the transceiver can select between 8 possible movements (2 horizontal, 2 vertical, and 4 diagonal). The intention is to sort these possible movements by localization precision priority, and then perform the feasible movement with highest priority. This is done in 3 steps: I. Compare separately confidence indicators of horizontal (north and south wall) and vertical walls (east and west): W0NS = sort(W0N ,W0S ) (7) W0EW = sort(W0E ,W0W ) II. Locate the wall with lowest energy: (8) W0NSWE = sort(W0NS (1),W0EW (1)) III. If the wall with lowest energy is horizontal one ( W0NSWE (1) equals W0NS (1) ) give more priority to vertical movement m1 = (W0EW (1),W0NS (1))

m2 = (0,W0NS (1)) m3 = (W0EW (2),W0NS (1)) m4 = (W0EW (1), 0) m5 = (W0EW (2), 0)

(9)

m6 = (W0EW (1),W0NS (2)) m7 = (0,W0NS (2)) m8 = (W0EW (2),W0NS (2)) Otherwise, more attention is given to horizontal exploration: m1 = (W0EW (1),W0NS (1)) m2 = (W0EW (1), 0) m3 = (W0EW (1),W0NS (2)) m4 = (0,W0NS (1)) m5 = (0,W0NS (2))

(10)

m6 = (W0EW (2),W0NS (1)) m7 = (W0EW (2), 0) m8 = (W0EW (2),W0NS (2)) with mi = (x, y) defined as movement vector of i-th priority, and composed of horizontal movement in x, and vertical movement in y. With this mobility model, a diagonal movement towards the intersection of two orthogonal walls with lower confidence indicators is preferred most. If this movement is not possible priority is given to movement towards the wall with lowest energy. IV. SYNTHETIC TEST ENVIRONMENT The synthetic test environment of the ICT-WHERE2 project [15] is a typical indoor environment (SIRADEL premises in France) as shown in Figure 2. This environment has been used to collect narrow-band and UWB measurements, which have been complemented by 3D Ray Tracing

Simulations using Sigint Solutions Radio Planning Tool (TruNET). TruNET is a full 3D Deterministic Ray Tracing (RT) simulator based on the Uniform Theory of Diffraction (UTD).

(on the left wall) and two wooden tables. The room is not perfectly rectangular with approximate size of 5.1mx7.3m and 2.4m height. The transceivers positions inside the room are equally spaced (118 points) at 1.5m height. The emitting power of transmitter was set to 10dBm, while the receiver sensitivity was set to -110dB. V. SIMULATION RESULTS

Fig. 2. The WHERE-2 synthetic environment

The transmitting antennas at 3.5GHz are assumed to be isotropic, whereas the received CIR have been calculated assuming both isotropic and also 8-sector (45°) horizontal aperture antennas. Out of these CIRs, different location/context-radio parameters, such as Received Signal Strength Indicators (RSSIs), multipath Times of Arrival TOAs, (Average) Power delay Profiles ((A)PDPs), channel delay spread, or mean excess delay, etc., can be extracted. The RT simulations take into account an unlimited number of reflections and refractions and up to one UTD diffraction. In order to maximize the accuracy of the RT simulations the electrical parameters of the building’s walls have been calibrated in the scope of WHERE-1 using in-situ measurements at 3.5GHz [16].

The simulation results correspond to 100 executions with transceiver set at arbitrary initial point and implementing the mobility model based on confidence indicator for controlling the movement trajectory. The results obtained after 2,3,4,5 and 10 observations, are presented. When analyzing the synthetic environment we found that the rays originated as rebounds from ceiling and floor have a great influence not only on omnidirectional CIR but also on just one sectorized antenna CIR, the one facing east. This is in conflict with the horizontal aperture requirement of the receivers’ sectorized antennas and not surprisingly produces erroneous east wall estimations. The ceiling/floor artifacts are cancelled, as seen in eq. (11), by changing the environment data of the first three samples of sectorized antenna facing east (corresponding to timestamps of 0ns,10ns and 20ns) CIR E ( 0ns ) = CIR E ( 0ns ) − 57.3243 [dB]

CIR E (10ns) = CIR E (10ns) − 46.6233 [dB]

(11)

CIR ( 20ns) = CIR ( 20ns) − 32.1402 [dB] E

E

The correct values are obtained by observing the CIR of the points that are more than 4m away from the east wall. If vertical reflection was not included all these positions would experience noise at first 3 samples (since there is no possible origin of horizontal reflection).

Fig. 3. The WHERE-2 synthetic environment Fig. 4. Environment exploration mobility model

In this paper, high-spatial resolution (0.5m) narrow-band (100MHz) RT simulations have been carried out in a single room of the synthetic environment as shown in Figure 3. The test room is a reception hall with four doors (2 at upper left corner, one at upper right corner, one at the center of the bottom wall). Its furniture is composed of metallic bookcase

The performance of exploration mobility model based on confidence indicator is seen in Figure 4, where 4 consecutive steps, together with real and estimated room map are given. The upper left image shows the estimated room dimensions after two observations. It can be observed that north end west

walls are estimated from the beginning with higher precision than south and east one, so the transceiver is moving in southeast direction. Also, since there is a metallic bookcase on the east wall (just bellow the initial point), the origin of these reflections is taken as as south wall estimations. It is for this reason that the south wall is the one with worst estimation. Eventually 3 additional steps are required to correct this first assumption. We should highlight the fact that even though transceiver was located between furniture (on south side) and wall (on the north side) the exploration of environment was always towards south. This is due to difference in power level of RTT readings related to furniture and wall reflections. Similar conclusion can be extracted when comparing west and east walls. Since the transceiver is closer to the west wall, the RTT reading consists of larger power levels, and hence its confidence indicator will be higher. Consequently, the transceiver will favor a move towards the east wall.

the number of observations is increased. For example, if the trajectory is limited to 2 steps, the error is between 20% and 50%, while for 10 observations is between 10% and 25%. Beside this advantage, the relative error in 90% of cases is reduced approximately to the half, from 25% to 12%, if the number of observation points is incremented from 2 to 10.

Fig. 6. CDF of average absolute positioning error as a function of number of observations

Fig. 5. CDF of wall dimension estimation relative error as a function of number of observations

The CDF of relative estimation error of room wall dimensions is given in Fig. 5. This relative error is directly related to the positioning error since transceiver position is provided as estimated distance with respect to the four room walls. It can be observed that CDF curve has two slopes: the greater one for CDF value of up to 0.8, 0.9 and the lower one for larger values. This property can be observed for all curves indicating the existence of precision limit of the approach regardless of the number of observations. The existence of this limit is easily noticed if we consider a transceiver located closely to the wall (less than 0.5m). In this case, the reflection originated at the wall will be of large strength and it’s peak will be inside the first time sample. According to the eq. (1) we would have two estimations, pointing to 0m and 1.5m. Since the weight of these estimations is very high the final wall estimation would be close to 0.75m, which compared to true distance value would suppose relative error of almost 50%. Nevertheless, the gain of larger number of observation points is seen in the low-slope region, which gets smaller if

The CDF distribution of the average absolute positioning errors is shown in Fig. 6. The average positioning error is defined as the average error of 4 walls distances estimations. This result shows that the average error in the worst estimation scenario, in meters, is reduced from 1.3m to 0.85m when using 10 instead of 2 observations. On the other hand, it can be observed that there is close to 50% of simulated scenarios in which incrementing the number of observation points, the localization precision is not improved, as the gain between 10 and 2 points is very low, from 0.6m to 0.5m. Again, this is due bandwidth limitations and due to number of observations close to the walls that produce erroneous estimations with overconfidence. Regardless of these limitations, the localization performance is improved with higher number of observation points. VI. CONCLUSION The algorithm for anchor-less self-localization and 2D mapping of rectangular rooms, based on CIR spatial discrimination and 100MHz narrowband signal, is presented. The spatial discrimination of the received signal is possible due to sectorized multiple receiver antennas with the total coverage of 360°. The novel component includes the introduction of instantaneous received power of the CIR’s strongest sample as indicator of position estimate quality. Hence wall distance estimate is not found based on simple mean average but on the weighed average. Furthermore, a sum of all confidence indicators corresponding to the same wall is proposed for measurement of wall precision estimation. This parameter is used as fundamental part of the recommended mobility model

developed with the aim of efficiently exploring the unknown environment. The algorithm is tested using synthetic data obtained with deterministic ray-tracing simulations of CIRs. It is shown that the even with 100MHz signal (up to 3m localization error per estimation due to sampling rate) the relative wall dimension estimation error is maintained in more than 80% of points below 9% and 12% for 10 and 5 observations respectively. The absolute positioning error with respect do real wall distance is below 0.6m in more than 70% of calculations obtained with 10 observations, while 0.8m error is obtained in the same percentage for only 2 observations. Beside confirming that this approach can be used for indoor localization and mapping, the analysis of the results affirms that better results are obtained with higher number of observations. Further work includes analysis of localization precision as a function of used bandwidth and horizontal aperture of sectorized antennas. Additional research to estimate number of walls in a non-rectangular room is also marked as valuable for generic SLAM solution. ACKNOWLEDGMENT This work has been carried out in the frame of the WHERE2 (FP7-ICT 248894) project, which is partly funded by the European Union and in the frame of Spanish MCIN project TEC2009-14219-C03-01. REFERENCES [1] J.J. Leonard and H.F. Durrant-Whyte. Mobile robot localization by tracking geometric beacons. Robotics and Automation, IEEE Transactions on, 7(3):376 –382, jun 1991. [2] R. Smith, M. Self, and P. Cheeseman. Estimating uncertain spatial relationships in robotics. In Robotics and Automation. Proceedings. 1987 IEEE International Conference on, volume 4, page 850, mar 1987. [3] H. Durrant-Whyte and T. Bailey. Simultaneous localization and mapping: part i. Robotics & Automation Magazine, IEEE, 13(2):99–110, 2006. [4] T. Bailey and H. Durrant-Whyte. Simultaneous localization and mapping (slam): Part ii. Robotics & Automation Magazine, IEEE, 13(3):108–117, 2006.

[5] Z. Sahinoglu, S. Gezici, and I. Guvenc, ”Ultra-wideband Positioning Systems: Theoretical Limits, Ranging Algorithms, and Protocols”, Cambridge University Press, 2008. [6] Filer N. P. Barton S. K. Guo W., Thomson S. L. Knowledge base assisted mapping for an impulse radio indoor locationsensing technique. In International Workshop on Wireless Adhoc Networks 2005, London, may 2005. [7] J. Youssef, B. Denis, C. Godin, and S. Lesecq, ”Dynamic IRUWB Channel Model Preserving IEEE 802.15.4a Statistics over Large-Scale Trajectories”, in Proc. IEEE PIMRC’09, pp. 17371741, Tokyo, Sept. 2009. [8] Wenyu Guo; Filer, N.P.; , "On the Accuracy of an Indoor Location-sensing Technique Suitable for Impulse Radio Networks," IEEE International Conference on Communications, 2007. ICC '07, pp.3987-3992, 24-28 June 2007 [9] P. Meissner, D. Arnitz, T. Gigl, and K. Witrisal, ”Analysis of an Indoor UWB Channel for Multipath-Aided Localization”, in Proc. IEEE ICUWB’11, pp. 565-569, Bologna, Sept. 2011. [10] Barton S. K. Guo W., Filer N. P. 2d indoor mapping and location-sensing using an impulse radio network. In IEEE ICU’05, Zurich, pages 296–361, Sept 2005. [11] Barton S. K. Guo W., Filer N. P. A novel wireless mapping and positioning technique for impulse radio networks. In 18th triennial URSI International Symposium on Electromagnetic Theory, Pisa 2004, volume 2, pages 712–714, may 2004. [12] Filer N. P. Guo W. 2.5d indoor mapping and location sensing using an impulse radio network. In IEE Seminar on Ultra Wideband Systems, Technologies and Applications 2006, London, pages 211–215, April 2006. [13] I. Arambasic, J. Casajus Quiros and I. Raos. Rectangular Room Dimensions Estimation Using Narrowband Signal and Sectorized Antennas, 5th International Symposium on Communications, Control and Signal Processing, Rome, Italy, May 2012. [14] I. Arambasic, M. Raspopoulos, J.C. Quiros, I. Raos, S. Stavrou, "Self-positioning and mapping of rectangular rooms with sectorized narrowband antennas," 20th International Conference on Software, Telecommunications and Computer Networks (Sept. 2012 [15] http://www.ict-where2.eu/. [16] “Measurements of location-dependent channel features," technical report, Deliverable D4.1 of the WHERE1 Project (ICT-217033), Oct. 2008.

ICT–248894 WHERE2

A.11

D2.6

Joint Anchor-less Tracking and Room Dimensions Estimation through IR-UWB Peer-to-peer Communications

V. La Tosa, B. Denis, and B. Uguen. Joint Anchor-less Tracking and Room Dimensions Estimation through IR-UWB Peer-to-peer Communications. In Proc. of The International Conference on Ultra Wideband 2011 (ICUWB’11), Bologna, Italy, Sept. 2011. c

2011 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

108 / 146

2011 IEEE International Conference on Ultra-Wideband (ICUWB)

Joint Anchor-less Tracking and Room Dimensions Estimation through IR-UWB Peer-to-peer Communications V. La Tosa, B. Denis

B. Uguen

CEA-Leti Minatec Campus 17 rue des Martyrs, 38054 Grenoble Cedex 09, France E-mails: [email protected], [email protected]

Universit´e Rennes 1 / IETR - UMR 6164 Av. du G´en´eral Leclerc, Campus de Beaulieu, 35042 Rennes Cedex, France E-mail: [email protected]

Abstract—In this paper we jointly address the problems of node location tracking and indoor mapping in the absence of infrastructure, relying on peer-to-peer IEEE 802.15.4a Impulse Radio - Ultra Wideband (IR-UWB) links. Coarse Time of Arrival (ToA) measurements, which are obtained within a low complexity Energy Detection (ED) receiver, are mapped into the direct path and single-bounce reflected multipath components. Then these measurements feed a multi-hypothesis tracking filter to retrieve the peers’ normalized coordinates and the room dimensions. Preliminary results are provided in a canonical mobile scenario, where dense multipath channels are generated through realistic UWB Ray-Tracing simulations. On this occasion, remaining structural transmitter/receiver and mirror ambiguities are pointed out and illustrated. Index Terms—Energy Detection, Extended Kalman Filter, IEEE 802.15.4a Standard, Impulse Radio, Indoor Environments, Simultaneous Localization and Mapping, Ultra Wideband, Time of Arrival, Tracking, Wireless Sensor Networks.

I. I NTRODUCTION A variety of applications would require simultaneously wireless communications, opportunistic self-localization and environment characterization capabilities at distributed mobile agents. One can cite for instance context-aware network optimization, robotics, house automation, self-reconfigurable Hi-Fi sound gears, augmented reality and gaming, assisted navigation in dangerous or confined environments. To fulfil these new requirements, one possibility is to rely on mobile ad hoc Low Data Rate (LDR) Wireless Sensors Networks (WSN), which can be opportunistically deployed in infrastructure-less environments. For such networks, indoor characterization mainly consists in retrieving geometric features of the environment (e.g. building layout, room volume). As for anchor-less localization, it refers to the ability for radio devices to infer their own coordinates in a local relative Cartesian system associated with the environment, without any fixed reference node. Ideally, the previous functionalities should be enabled (almost) for free while communicating, out of transmitted radio signals. In this context, the fine temporal resolution and the flexibility of the Impulse Radio - Ultra Wide Band (IR-UWB) technology seem particularly appealing. In one first solution, depicted as Indoor Mapping [1], pairs of communicating nodes interpret the received channel impulse response as a temporal fingerprint that accounts for the geometrical configuration around, delivering a four-wall map of the room. However, nodes mobility and realistic Time of Arrival (ToA) measurement errors have not been really considered. A second Imaging approach

978-1-4577-1762-8/11/$26.00 ©2011 IEEE

[2] inherits from radar techniques and benefits from nodes mobility to infer the presence of scatterers and walls through cross-correlations. Antenna arrays are required on both transmitting and receiving sides, and hence, integrating this solution in an opportunistic WSN context looks problematic. More recently, original works reported in [3] considered using a battype bistatic UWB radar for autonomous indoor navigation in the presence of unknown obstacles. The system is composed of one transmitter and two receivers arranged in a fixed linear array, recalling the biological sonar used by bats and cetaceans. The environment is reconstructed out of impulse responses by extracting features like walls, edges and corners, according to the set of prior canonical propagation models. These identified back-scatterering elements are then used as virtual landmarks to navigate, but this technique does not include communication means. Finally, another approach was put forward in [4] for joint static localization and room dimensions estimation. In this solution, which relies on a deterministic electromagnetic model, cross-correlations are performed between local templates and received IR-UWB multipath components to determine Angles of Incidence (AoI) to reflecting surfaces and ToAs, or even Angles of Arrival/Departure (AOA/D). Following a Maximum Averaged Likelihood (MAL) approach, the measurements are defined as functions of the relative node coordinates and room dimensions. The latter are finally estimated through non-linear optimization, while progressing in an estimation tree as new paths are detected. Overall, those investigations enabled to point out numerous remaining challenges from both hardware and computational points of view, even under static and simplified channel assumptions. In this paper, coarse energy-based ToA estimates associated with direct and single-bounce reflected paths (Section II) are provided as observations into a multi-hypothesis Extended Kalman Filter (EKF), offering a linearized approximated solution to our non-linear joint mapping and localization problem (Section III). First, in comparison with [4], a more realistic channel model is considered under nodes mobility, authorizing higher order electromagnetic interactions in our simulations. Moreover, alleviated computational and hardware capabilities are now required on the radio device side, assuming a simple Energy Detection (ED) receiver architecture. Further system constraints are set to partly comply with the IEEE 802.15.4a standard [5], which explicitly addresses ranging-enabled, lowcost and low-power WSN applications based on IR-UWB. Mobility is expected to compensate for losses in the received signal dynamics e.g., providing spatial correlation and mea-

560

surements redundancy. Mobility and low-complexity features would also favor the claimed opportunistic applications. As a first attempt in the process of solving this particularly complex problem, only 4-wall rectangular rooms are considered herein and performance is assessed (Section IV) through one single illustrating setting (i.e. particular trajectories in one room), what is obviously restrictive and still significantly simplified. However, it is expected that the proposed technique could be gradually extended into more realistic scenarios (e.g. with nonregular rooms, pieces of furniture, NLOS...), by incorporating for instance further measurements like the ToAs of double reflected paths (for which simple trigonometric equations can still be written).

number of possible measurements’ quadruplets to be tested, authorizing more flexibility in the detection process (e.g. coping with multipath synchronization losses, overlap effects under poor time resolution, crossing paths under mobility). Note that Nmeb can be even larger if several bins have the same energy levels for the 6 − th path selection step.

II. M ULTIPATH T OA E STIMATION WITH ED R ECEIVERS Stringent technological constraints are usually imposed to LDR radio solutions covering WSN applications, especially in terms of power consumption and complexity. In this context, simple non-coherent ED receivers have been favored for the last past years [6], [7]. A typical receiver architecture is represented on Figure 1, including a UWB antenna, a Band Pass Filter (BPF) filtering in the transmission band BIEEE some Additive White Gaussian Noise (AWGN) with bilateral Power Spectrum Density (PSD) equal to N0 /2, a Low Noise Amplifier (LNA) with a gain GLN A , a signal squaring block, an integration stage over small time intervals of duration ∆s , and a sampling block running at the frequency fs = ∆1s . Classically, the Noise Figure N F can account for further imperfections in the receiving RF chain. The energy samples collected in successive time bins are quantized over Nbits through an Analog to Digital Converter (ADC) with a saturation voltage VADC . Finally, the quantized energy values are summed up at each time bin over transmitted sequences within a baseband accumulator block (ACC). Figure 2 shows

Fig. 1.

Energy Detection (ED) receiver block diagram

an example of received analog signal (a) generated through Ray-Tracing [8] with the corresponding accumulated ED result (b), in a 6m×5m rectangular room, with N0 = −174dBm/Hz, N F = 10dB, BIEEE according to the IEEE 802.15.4a channel #4, GLN A = 25dB, fs = 250MHz, Nbits = 4 bits, VADC = 0.9V. Strong degradations can be noticed in the received signal representation (i.e. due to sub-Nyquist frequency sampling and limited signal dynamics), what is inherent to the chosen receiver architecture. In our proposal however, the ToAs of the most significant multipath components (assumed to result either from the direct path or from a single-bounce reflection) must be determined out of the accumulated energy samples. The retained measurements are based on the Neb non-null most energetic ED bins. If by chance Neb = 5 (e.g. due to a particular ADC threshold setting and sparse estimated energy profiles), the earliest selected time bin is systematically associated with the estimated direct path ToA τˆ(0) and the 4 remaining bins are associated with the ToAs of the assumed single-bounce multipath components. When Neb > 5, at least the Nmeb = 6 most energetic bins are selected. Here, Nmeb impacts the

Fig. 2. Example of received signal voltage after a LNA with GLN A = 25dB, N F = 10dB (a) (simulated through UWB Ray-Tracing), and accumulation result over 32 consecutive PRPs (b) of energy samples quantized on Nbits = 4bits (under the saturation voltage VADC = 0.9V).

III. EKF- BASED J OINT T RACKING AND ROOM D IMENSIONS E STIMATION Beside hardware capabilities, low-complexity and lowpower requirements impose the use of simple embedded estimation algorithms e.g., limiting the number of non-linear operations. Thus, an Extended Kalman Filter (EKF), whose recursive linearized structure fits well to real-time constraints, is fed here by the ED-based ToA measurements introduced in Section II. A. State Vector and Equation The state variables to be estimated shall enable joint anchorless localization and environment characterization. Accordingly, the state vector x(k) at time step k includes the normalized Cartesian coordinates of the transmitter and receiver nodes, respectively x˜t (k) , y˜t (k) , x˜r (k) and y˜r (k) , the room (k) dimension Dx along the x axis, and finally the room di(k) (k) (k) mensions ratio RD = Dx /Dy , as follows: h iT (k) x(k) = x˜t (k) y˜t (k) x˜r (k) y˜r (k) Dx(k) RD (1) Note that the coordinates are normalized with respect to the room dimensions, i.e. x˜t , y˜t , x˜r , y˜r ∈]0, 1[, while the actual coordinates xt , xr ∈]0, Dx [ and yt , yr ∈]0, Dy [. This formalism provides an explicit dependency of ToAs on environment variables Dx and RD , what makes their estimation easier. As an example, referring to Figure 3, the ToA τ0 of the direct path can be straightforwardly written as: q 2 2 (2) τ0 = 1c Dx (x˜r − x˜t ) + (y˜rR−2y˜t ) . D

where c is the speed of light in the vacuum. The previous equation implicitly means that ToAs are not measured on a relative time scale but that the system also enables absolute timing and hence ranging (e.g. through Two Way Ranging transactions). The state transition equation from time step k − 1 to time step k, based on the latest state vector x(k−1) , is as follows:

561

x(k) = F(k) x(k−1) + w(k)

(3)

where the linear transformation matrix F(k) is simply defined as I6 and the state noise vector w(k) is centered Gaussian with a covariance matrix  2  σ∆xyt I2 0 ... 0  ..  2  0 σ∆xy I 0 .  (k)  r 2  (4) Q =  . .. 2  0 σDx 0  2 0 ... 0 σR D

Note that, according to the same state equation, the effects of actual 2D speeds (and potential accelerations) on actual positions are accounted directly within the coordinates uncertainty model, the latter being characterized by the variance 2 2 terms σ∆xy and σ∆xy in Q(k). In order to make the solution t r finite in space for the state vector x(k) , Dx and RD are assumed to lie in [Dxlb , Dxub ] and [RD lb , RD ub ] respectively. Furthermore, these variables should be theoretically constant in time if the nodes are confined into one room. Nevertheless, it is expected that the EKF starts from rough estimates, which are subsequently refined thanks to devices mobility. Therefore, in order to let the filter slightly tune the estimates for those two environmental parameters, Dx and RD are also considered as affected by centered Gaussian noise terms with small standard deviations σDx = (Dxub − Dxlb )/10  (Dxub − Dxlb ) and σRD = (RDub − RDlb )/10  (RDub − RDlb ) respectively. B. Observation Vector and Equation Assuming unbiased measurements, the observation vector z (k) at time step k can be related to the state vector x(k) as follows:   z (k) = h(k) x(k) + v (k) (5)

where h(k) (·) is a vector-valued non-linear function mapping the true state space into the observation space and v (k) is the Gaussian centered observation noise vector with a covariance matrix R(k) . As already mentioned, this observation vector z (k) at time (k) step k should be theoretically built with {ˆ τ(i) }i=0..4 , i.e. the set of ToA measurements associated with the detected ˆ direct path and presumed single-bounce reflected paths. (·) recalls that those ToAs are issued from a preliminary channel estimation algorithm based on ED and (·)(i) simply means that τˆ(i) > τˆ(0) , ∀i > 1. Two labeling systems coexist here. The path labeling mentioned above, which depends on the path detection order (and hence, on the energy of multipath components), shall not be confused with that introduced in Figure 3, which is related to an arbitrary wall numbering strategy (unknown at the receiver but set a priori in the filter room model). It will be understood later in Subsection III-C how the best measurements quadruplet is determined out of the Nmeb bins selected originally (See Section II) and how the wall-path mapping problem is solved. Whatever the path labeling and measurements down-selection strategies, we assume in the filter that five independent ToA measurements of equivalent quality are selected, so that R(k) is simply characterized by στ2 I5 , if στ2 is the a priori variance of unitary path ToA measurements. The nonlinear dependency of the true direct path ToA τ0 on state variables has already been displayed in (2). Relying on simple trigonometric relationships, the ToA τ1 and τ2 for the subsequent single-bounce paths (and similarly for τ3 and

Fig. 3. Simplified single-bounce reflections in the peer-to-peer indoor scenario, with arbitrary multipath labeling

τ4 ) can be written as: τ1

=

τ2

=

y˜ 1 1  r +y˜t  c Dx RD cos arctan R |x˜r −x˜t | D y˜r +y˜ t 2− 1  x˜r −x˜t  c Dx cos arctan 1 |y˜r −y˜t | R x˜r +x˜ −2 D

(6)

t

where the path label n in τn now corresponds to the actual wall number assumed in the filter, as defined in Figure 3. C. Filter Predictions and Updates The filter aims at delivering x ˆ (k|k) , which is the a posteriori state estimate at time step k, given the observation up to (and including) time step k. Classically, in a first prediction stage, the a priori state x ˆ (k−1|k−1) (i.e. the latest available state estimate) and the related covariance P(k−1|k−1) are respectively predicted into x ˆ (k|k−1) and P(k|k−1) , as follows: x ˆ (k|k−1) = F(k) x ˆ (k−1|k−1)  T P(k|k−1) = F(k) P(k−1|k−1) F(k) + Q(k)

(7)

Then, during the next correction stage, the following quantities will be computed:   ˆe(k) = z (k) − h(k) x ˆ (k|k−1)  T (8) S(k) = H(k) P(k|k−1) H(k) where ˆe(k) is the observation innovation, S(k) is the approximated covariance of the linearized model of ˆe(k) , and H(k) is the Jacobian of h(k) evaluated at x = x ˆ (k|k−1) :   ∂τ0 ∂τ0 ···  ∂ x˜t ∂RD   . ..  ..  . H(k)  (9) . .   .  ∂τ  ∂τ4 4 ··· ∂ x˜t ∂RD x=ˆx(k|k−1)

Then, the Kalman gain K(k) , the a posteriori state estimate x ˆ (k|k) and finally the a posteriori estimate covariance P(k|k) are calculated as follows:  T  −1 S(k) K(k) = P(k|k−1) H(k) (k) (k) (10) x ˆ (k|k)= x ˆ (k|k−1) + K ˆe P(k|k) = I10 − K(k) H(k) P(k|k−1)

But the filter must still deal with the unknown mapping between the measured ToAs and the true ToAs numbered according the walls labeling of Figure 3. In Line of Sight (LOS) scenarios at least, what is an a priori assumption here (detectable based on ED through e.g., [7]), the first measurement can be systematically associated with the true

562

direct path ToA (i.e. τˆ(0) always mapped onto τ0 ). For the four remaining ToAs, 4! = 24 arrangements are possible, that is to say 24 possible vectors z (k) . Inspired by [9], the optimal measurements set is chosen according to the decision variable (k) ξe , which is the normalized scalar innovation:  T  −1 ξe(k) = ˆe(k) S(k) ˆe(k) (11) (k)

The Gaussian assumption for z (k) implies that ξe is a central Chi-Square distributed random variable with dim{ˆe(k) } = 5 degrees of freedom and unitary variance. Then, among the 24 possible observations (i.e. ToAs arrangement), the one whose (k) ξe shows the highest likelihood with respect to the ChiSquare PDF is chosen at each time step k. Note that the ToAs order may change under nodes mobility but the proposed criterion still ensures that the most likely arrangement is always preferred at each instant. Moreover, among the Nmeb selected bins, no more than 5 measurements practically correspond to the searched ToAs. Three main hypotheses are considered: • H0 : there are no overlapping paths, so that the earliest bin is associated with τˆ(0) and the 3 following most energetic bins with 3 single-bounce paths, but the Nmeb − 4 remaining bins are kept open for the 4 − th and last ToA; • H1 : 1 path overlaps with another path in 1 of the 4 most energetic bins; • H2 : 2 paths overlap with 2 other paths (i.e. 2 overlapping paths in each of the 2 bins). For the three hypotheses, all the possible multipath combinations are then evaluated, given the decision variable in (11). The combination leading to the highest conditional likelihood is chosen (and so is the measurements quadruplet accordingly). Finally, according to [10], the known spatial bounds for state variables (i.e. x˜t , y˜t , x˜r , y˜r ∈]0, 1[, Dx ∈ [Dxlb , Dxub ] and RD ∈ [RD lb , RD ub ]) enable to guide the EKF estimation ˆ k|k showing outprocess. Predictions x ˆ k|k−1 and corrections x of-bound values are forced to stick within the closest upper or lower bound. IV. S IMULATIONS AND R ESULTS A. Tested Scenario and Simulation Set-up 1) Assumed Refreshment Period and Signal Parameters: In our simulations, further IEEE 802.15.4a [5] constraints were considered, impacting both the ToA measurements refreshment rate and received signal shape. Following the standard band plan, the fourth channel of the low band was selected, with a central frequency fc = 3.994GHz and a bandwidth at −3dB of the maximum PSD equal to BW−3dB = 1.331GHz. Beside, the Middle Alva antenna used in [11] was integrated in our UWB Ray-Tracing simulations. Its low cost and light weight make this antenna suitable for realistic WSN applications. The Rx sampling frequency was equal to fs = 250MHz. Finally, the PSD of the simulated signal was compliant with the FCC regulation mask, with a PRP of 62ns (i.e. the shortest available in the standard). In addition, the structure of a typical IEEE 802.15.4a PHY Protocol Data Unit (PPDU) was respected, where the preamble used for synchronization and ranging was 64-symbol long. As for the other fields, the Start of Frame Delimiter (SFD) was 8-symbol long, the PHY HeadeR (PHR) comprised 19 bits (including the Ranging Packet bit), and finally 32 bytes were devoted to the PHY Service Data Unit (PSDU). Overall, under these

packet specifications, one could achieve a bit rate of 850 kbps. Considering an integer multiple of the frame duration, the refreshment period tupdate between consecutive ToA measurements (and hence, between consecutive tracking results) was equal to 55.6ms. Finally, in order to save complexity, during the ToA estimation phase, it was assumed that only 50% of the preamble symbols could be accumulated before issuing one accumulation result (i.e. one set of multipath ToA measurements). 2) Mobile Scenario: A canonical mobile scenario in a 6m×5m room was simulated and tested, offering changing configurations in terms of multipath overlap and relative geometry. The actual mobility model was characterized by an initial speed v0 for both x and y components and an acceleration a affecting arbitrarily the y component only. The simulated dynamic nodes coordinates (x(t), y(t)) could then be fully expressed as two deterministic functions of the elapsed time t, as follows: x(t) = x0 + v0,x t y(t) = y0 + v0,y t + 21 at2

(12)

with v0,x,tx = −0.05ms−1 , v0,x,rx = 0.55ms−1 , v0,y,tx = −0.035ms−1 , v0,y,rx = 0.15ms−1 , atx = 0.005ms−2 , arx = 0.05ms−2 in our simulations. In the filter, the transmitter nodes was also assumed to be slower than the receiver node, with the respective speed norms vslow ∈ [vslowlb , vslowub ] = [0, 0.1]ms−1 and vf ast ∈ [vf ast lb , vf ast ub ] = [0, 1]ms−1 . Accordingly, the coordinates uncertainty was also adjusted: σ∆xyt = σ∆xyslow = σ∆xyr = σ∆xyf ast =

RD ub vslow ub Dx lb 3 RD ub vf ast ub Dxlb 3

(13)

making the state equation homogeneous and the state error model consistent with the actual mobility pattern. In addition, we consider the upper and lower intervals [Dxlb , Dxub ] = [1, 25]m and [RD lb , RD ub ] = [1, 2] for bounding room dimensions in the filter. These figures have been chosen according to a survey, which was carried out over seven typical buildings belonging to the CEA-Leti premises (Grenoble, France) and comprising about 150 rooms (i.e. exempting corridors). Finally, assessing mostly the tracking performance in a steady-state regime (i.e. once the target is locked), the initial condition x ˆ0 was obtained as: x ˆ 0 = x0 + w 0

(0)

(14)

(0)

where w0 follows the same distribution as that of w(k) in (3) so that P0|0 was equal to Q(k) . B. Simulation Results To characterize estimation as a function of the elapsed time under mobility and to illustrate possible multimodal effects, at each time step, an empirical PDF of the estimated state variables is calculated over noise draws (i.e. over different randomly noised received signals at each occupied locations along one single trajectory). For instance, the PDFs of the estimated coordinates are shown on Figure 4, namely Tx abscissa (left top) and ordinate (left bottom) and Rx abscissa (right top) and ordinate (right bottom). For each coordinate, the true dynamic value is represented by a filled red square, while the mirror value (resulting from flip ambiguities as pointed out in [12]), is a filled yellow triangle. Coordinates

563

tracking looks challenging for the receiver (i.e. for the fastest node), whereas better results are obtained for the transmitter. Beside, Figure 5 shows similar results for Dx and RD . In

Fig. 6. Empirical PDFs of estimated Tx (left) and Rx (right) normalized coordinates, as a function of the time step, with unbiased TOA measurements

Fig. 4. Empirical PDFs (over noise realizations) of estimated Tx (left) and Rx (right) normalized coordinates, as a function of the time step

Fig. 5. Empirical PDFs (over noise realizations) of estimated room main dimension Dx and room dimensions ration RD , as a function of time step

the filter model, {z (k) } in (5) are assumed to be affected by centered Gaussian noise terms, what is often unverified with real ToA measurements. Hence, to assess the impact of such biases, unbiased {z (k) } have also been considered in the same scenario. On Figure 6, signs of ambiguity are then present for the estimated Tx ordinate (bottom left). The PDFs are split into two distinct modes, one of them being centered around the true value. Surprisingly, the other mode does not fluctuate around the expected mirror values (i.e. yellow triangles), but around the mirror values of the true Rx ordinate (bottom right). Therefore, joint Tx/Rx and mirror ambiguities affect the ordinates here, but less evidently the abscissae. This asymmetry is likely due to the very next initial value for ordinates, in comparison with the other node mirror images shown by yellow triangles. Overall, if the four traces corresponding to all the possible estimates were superposed, one combination would clearly enable proper tracking. Despite ambiguities, other simulations showed that Dx and RD are well estimated, with smaller fluctuations than on Fig. 5. V. C ONCLUSION We have presented a solution for anchor-less tracking and room mapping trough peer-to-peer IR-UWB links. Considering low-complexity and IEEE 802.15.4a constraints, energybased ToA measurements are associated with direct and singlereflection paths to feed an EKF tracking filter. Multiple hypotheses are then tested to handle continuous mapping

between detected paths and reflecting walls under mobility. Rather satisfactory simulation results could be obtained in the tracking steady-state regime, especially under unbiased measurements. But combined Tx/Rx and mirror ambiguities have also been pointed out, disclosing new axes of investigation (e.g. aiming at a univocal mapping between multipath components and walls). Future works could also address NLOS, cooperative mapping (with > 2 nodes), non-rectangular rooms and comparisons with other baseline methods. ACKNOWLEDGMENTS This work has been carried out in the frame of the eUWB (FP7-ICT 215669) and WHERE2 (FP7-ICT 248894) projects, which are partly funded by the European Union. R EFERENCES [1] W. Guo and N. P. Filer. On the accuracy of an indoor location-sensing technique suitable for impulse radio networks. In Proc. IEEE ICC ’07, Glasgow, UK, June 2007. [2] R. Zetik, J. Sachs, and R. Thoma. Imaging of propagation environment by UWB channel sounding. Technical report, COST 273 TD(05) 058, January 2005. [3] T. Deissler and J. Thielecke. Feature based indoor mapping using a battype uwb radar. In Proc. IEEE ICUWB’09, pages 475–479, Vancouver, Canada, September 2009. [4] V. La Tosa, B. Denis, and B. Uguen. Impact of antennas on the anchor-less indoor localization of a static IR-UWB pair. In Proc. IEEE PIMRC’10, Istanbul, Turkey, September 2010. [5] IEEE 802.15 Task Group 4. Part 15.4: Wireless medium access control (MAC) and physical layer (PHY) specifications for low-rate wireless personal area networks (WPANs). IEEE Std 802.15.4a-2007 (Amendment to IEEE Std 802.15.4-2006), 2007. [6] I. Guvenc, Z. Sahinoglu, and P.V. Orlik. Toa estimation for ir-uwb systems with different transceiver types. IEEE Trans. on Microwave Theory and Techniques, 54(4):1876 – 1886, june 2006. [7] A. Rabbachin, I. Oppermann, and B. Denis. Ml time-of-arrival estimation based on low complexity uwb energy detection. In Proc. IEEE ICUWB’06, pages 599 – 604, Waltham, USA, sept. 2006. [8] F. Tchoffo-Talom, B. Uguen, E. Plouhinec, and G. Chassay. A sitespecific tool for UWB channel modeling. In Proc. Joint UWBST & IWUWBS, pages 61–65, Kyoto, Japan, May 2004. [9] S. Sukkarieh, E. M. Nebot, and H. F. Durrant-Whyte. A high integrity imu/gps navigation loop for autonomous land vehicle applications. IEEE Trans. on Robotics and Automation, 15(3):572–578, 1999. [10] R. Kandepu, L. Imsland, and B.A. Foss. Constrained state estimation using the unscented kalman filter. pages 1453–1458, Ajaccio, France, June 2008. [11] R. D’Errico and L. Ouvry. Time-variant BAN channel characterization. In Proc. IEEE PIMRC’09, Tokyo, Japan, September 2009. [12] V. La Tosa, B. Denis, and B. Uguen. Maximum averaged likelihood estimation tree for Anchor-Less localization exploiting IR-UWB multipaths. In Proc. IEEE VTC Spring 2010, Taipei, Taiwan, May 2010.

564

ICT–248894 WHERE2

A.12

D2.6

Benefits from Cooperation in Simultaneous Anchor-less Tracking and Room Mapping based on Impulse Radio - Ultra Wideband Devices

M. Des Noes and B. Denis. Benefits from Cooperation in Simultaneous Anchor-less Tracking and Room Mapping based on Impulse Radio - Ultra Wideband Devices. In Proc. of The International Conference on Systems, Signals, and Image Processing 2012 (IWSSIP’12), Special Session on Localization and Tracking, Vienna, Austria, April 2012. c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

114 / 146

Benefits from Cooperation in Simultaneous Anchor-less Tracking and Room Mapping based on Impulse Radio - Ultra Wideband Devices M. Des Noes, B. Denis CEA-Leti Minatec Campus 17 rue des Martyrs, 38054 Grenoble Cedex 09, France E-mails: [mathieu.desnoes, benoit.denis]@cea.fr Abstract—In this article, a method is proposed to jointly estimate both the dimensions of a rectangular room and the relative locations of wireless devices communicating in this room. Relying on Low Data Rate (LDR) Impulse Radio - Ultra Wideband (IR-UWB) links, one can determine accurate estimates of the propagation delays corresponding to the direct path and the four single-bounce paths reflected on surrounding walls. Once properly ordered over each single link, these five delays enjoy explicit mathematical relationships with the transmitter/receiver nodes coordinates and the dimensions of the rectangular room. We propose herein new solutions to related underlying problems such as data association, initialization and tracking through Extended Kalman Filter (EKF) in a cooperative context including more than two mobile devices. The impact of the Time Of Arrival (TOA) estimation precision, the initial guess and the actual room dimensions is then evaluated through canonical simulation scenarios. Index Terms—Extended Kalman Filter, Impulse Radio, Localization, Mapping, Wireless Sensor Networks, Ultra Wideband.

I. I NTRODUCTION Many emerging applications such as context-aware network optimization, self-reconfigurable Hi-Fi systems, augmented reality in gaming or assisted navigation in confined environments, require that mobile wireless nodes perform simultaneously self-localization and environment mapping while communicating with limited access to elements of infrastructure. For this sake, one can take benefits from the multipath propagation delays classically observed over radio links in indoor environments. For instance, considering the emerging Low Data Rate (LDR) Impulse Radio - Ultra Wideband (IRUWB) technology [1], which provides unprecedented resolution capabilities, it is theoretically possible to detect and distinguish accurately between the various received paths through standard Time of Arrival (TOA) estimation [2], [3]. Among all these received multipath components, the echoes issued from simple sequences of electromagnetic interactions (e.g. singlebounce reflected paths) are indeed assumed to be the most energetic (hence, more easily detectable and more systematically observable) and persistent under mobility due to spatial correlation effects [4]. The processing of the corresponding excess delays, which naturally convey information about the geometry of the involved radio links, hence makes possible the estimation of the room dimensions and the coordinates of the communicating nodes in this room. Several studies reported in the recent literature already attempted to address this problem. Unfortunately, either they adopt a bat-type UWB radar approach [5], [6] requiring a highly specific hardware configuration (i.e. central transmitter with a receive antennas array) or they require monostatic

channel impulse response measurements (i.e. self-to-self channel sounding) [7] that may be not supported in conventional communication-oriented contexts involving peer-to-peer Single Input Single Output (SISO) communications. In this paper, a new method is proposed to jointly and simultaneously estimate both the dimensions of a rectangular room (i.e. mapping) and the relative locations of communicating devices in this room (i.e. relative positioning), coping with various pairs of mobile devices in a coordinated and cooperative context. Though exclusively restricted to rectangular rooms so far (which is anyway a rather common case within most buildings and, especially industrial factories or warehouses), the proposed scheme still claims to address a rather vast range of applications. In this study, the geometric information conveyed in the four delays originated by the single-bounce reflections is exploited similarly to the noncooperative solution in [8], relying on explicit mathematical relationships. However, we derived herein a different quadratic system model linking the squared value of the measured multipath delays with the unknown parameters associated with three cooperative mobile nodes. This model is more adequate to perform initialization through the minimization of quadratic cost functions (and the optional data association update) in a tracking Extended Kalman Filter (EKF) and it helps to understand the hidden problem symmetries. The paper is organized as follows. Section II describes the system model as an input to our algorithm definition detailed in Section III. Section IV illustrates the performance of the proposed solution in simulated canonical scenarios and Section V concludes the paper.

II. S YSTEM M ODEL

As shown on Figure 1, we first assume two nodes communicating in a rectangular room over Line Of Sight (LOS) radio links, including a direct path and 4 single-bounce reflections on surrounding walls. Standard 2D Cartesian coordinates (xm , yn ) and (xm , yn ) are used to represent the locations of nodes m and n, respectively. The constant dimensions of the room are noted Dx and Dy along the x and y axes respectively. At each discrete time instant kT (with T the refreshment period), the theoretical delays associated with the 5 multipath k components {τmn,j }j=0..4 mentioned above are related to the

6 unknown geometrical variables as follows:  k  2 k 2 k 2 k 2 k 2 T (τ = mn,0 ) (τmn,1 ) (τmn,2 ) (τmn,3 ) (τmn,4 ) (xm (k) − xn (k))2 + (ym (k) − yn (k))2   (xm (k) − xn (k))2 + (ym (k) + yn (k))2   1  2 2 (2Dx − xm (k) − xn (k)) + (ym (k) − yn (k))  c2   (xm (k) − xn (k))2 + (2Dy − ym (k) + yn (k))2  (xm (k) + xn (k))2 + (ym (k) − yn (k))2

(1) where c = 3.108 m/s is the celerity of light, n and m are the indexes of the 2 communicating nodes and where the reflecting surfaces are numbered according to an arbitrary order j = 0..4. In real-life systems, the multipath delay profile between each pair of devices may be estimated using a receiver compatible with the IEEE 802.15.4a standard [1]. For instance the delays associated with some of the strongest energy bins in a non-coherent energy detector might be considered [8]. As the multipath TOA procedure does not fall in the scope of our investigations here, we simply assume that 5 time measurements affected by Additive White Gaussian Noise 2 (AWGN) terms {bkmn }, which are i.i.d. with a variance σm for all the measurements, are squared into unordered observations {tkmn } as follows: k tkmn,i ←→ (τmn,j + bkmn,j )2 , i, j = 0..4

(2)

It is hence worth noting that the squared delays measured at the receiver are indexed by a variable i = 0..4, which may not coincide with the index j = 0..4 of the prior wall numbering model in Eq. (1). This is due to the fact that the receiver has no a priori knowledge of the correct mapping association.

through TOA estimation need to be mapped onto the a priori modeled {τj }j=0..4 of Eq. (1). In LOS channels, the first measured squared delay t0 (i.e. first being intended in terms of arrival time but not in terms of energy or detection order) can be associated with the direct path and hence systematically mapped onto τ0 . Then t1 can be mapped to τ1 , τ2 , τ3 , or τ4 , and so on for the other measured squared delays t2 to t4 . Overall, there are 24 remaining mapping combinations to be tested for each pair-wise peer-to-peer link. Besides, preliminary simulations carried out in a tracking EKF context showed that the results were very sensitive to with a random initialization of the state vector, even when assuming perfect data association. A theoretical explanation to this EKF filter instability problem under poor initialization with non-linear dynamics is given in [10]. Hence, it is required to provide a sufficiently reliable initial guess of the estimated parameters. The proposed solution consists in minimizing a quadratic cost function to provide both the initial parameters estimates and the initial correct mapping (the mapping being possibly revised afterwards, e.g. out of tracking filter innovation outputs). In order to keep the computation load of this initialization step acceptable, we limit the problem to L = 2 links involving N = 3 nodes. Let us for instance consider the links involving node 1 first (i.e. link (n = 1, m = 2) and (n = 1, m = 3)). From the 5 measurements performed over each link, one can build 24 (ordered) candidate vectors {z12,a }a=1..24 and 24 more vectors {z13,b }b=1..24 , representing the mapping combinations to be tested over the 2 links, e.g. z12,1 = [t12,0 t12,1 t12,2 t12,3 t12,4 ]T , z12,2 = [t12,0 t12,1 t12,2 t12,4 t12,3 ]T , etc. There is thus a total of 242 = 576 combinations. For each link and each built observation vector z12,a or z13,b , an error vector can be formed as follows: 12,a = z12,a − h(x1 , y1 , x2 , y2 , Dx , Dy ), a = 1..24 13,b = z13,b − h(x1 , y1 , x3 , y3 , Dx , Dy ), b = 1..24

(3)

The function h(.) can be drawn from Eq. (1). A quadratic cost function involving node 1 as the local central node can thus be defined as follows: Fig. 1. Single-bounce reflection model for 2 nodes communicating over Line of Sight links in a rectangular room, with an arbitrary wall numbering (*).

III. A LGORITHM D ESCRIPTION The previous data association problem can be solved either once for all during the initialization phase (i.e. considering the initially determined mapping can remain unchanged afterwards) and/or continuously updated in the steady-state tracking regime. Accordingly, at each time step k under mobility after initialization, the overall algorithm synopsis would be as follows: received multipath acquisition and delays estimation (not addressed herein), revised data association (optional), useful parameters estimation and continuous tracking through EKF filtering [9]. Hereafter, we will consider that the initialization procedure provides a definitive mapping. A. Initial State Estimation and Data Association For the purpose of alleviating the notations, we deliberately omit the time step k first, as well as the nodes indexes m and n. The 5 measured squared delays {ti }i=0..4 generated

g1,ab (x1 , y1 , x2 , y2 , x3 , y3 , Dx , Dy ) = ||12,a ||2 + ||13,b ||2 , ab = 1..576 (4) At the very beginning of our investigations, a first cost involving only 2 nodes used to be tested, with e.g. ga (x1 , y1 , x2 , y2 , Dx , Dy ) = ||a ||2 , a = 1..24. The performances were not satisfactory since the minimization operation does not converge towards a unique optimum. Actually, it seems that the set of solutions in such single-link cases may be a line or a surface, or even multiple discrete points. As a consequence, the new cost function involving 3 nodes was selected instead and the minimization procedure was shown to converge systematically towards a unique solution, hence illustrating the benefits from cooperative links. The positions and the room dimensions are thus the parameters that minimize the cost function in Eq. (4). Since it is convex, the global minimum can be obtained with standard linear programming techniques [11]. The initial state estimation requires performing this minimization operation for each of the 576 combinations. The selected set of parameters is the one among the 576 combinations giving the smallest error. Finally, the correct initial data association mapping (ˆ a, ˆb)

(between {ti }i=0..4 and {τj }j=0..4 ) and the initial condition ˆ 1 (0) are the ones that lead to this smallest error, according X to the following rule: ˆ 0 = argmin [g1,ab (X)], ab = 1..576 X X 1,ab ˆ 0 )] (ˆ a, ˆb) = argmin(a,b) [g1,ab (X 1,ab ˆ 1 (0) = X ˆ0 X ˆ 1,ˆ ab

(5)

More generally speaking, 24L combinations for L links in cooperative scenarios must be tested (i.e. with N > 2 nodes). But it is rather clear that only a sub-set of the N (N − 1) available links shall be used (i.e. non-exhaustive cooperation) to preserve low consumption, low traffic and low complexity. Even if the computational burden remains quite important, note that the initial calculation could anyway be externalized and centralized, the wireless nodes being only in charge of acquiring the received signals over cooperative links. The proposed solution does not always converge to the same mapping and coordinates representation. This is due to structural transmitter/receiver and mirror ambiguities. Eq. (1) helps us to determine these ambiguities. One can observe that: τ0 (xn , yn , xm , ym ) = τ0 (yn , xn , ym , xm ) τ2 (yn , xn , ym , xm , Dy ) = τ3 (xn , yn , xm , ym , Dx ) τ4 (yn , xn , ym , xm ) = τ1 (xn , yn , xm , ym )

(6)

Hence exchanging the x and y dimensions, with the appropriate mapping of the measured delays to the model, also lead to a valid result. Moreover, one also observes that: τ0 (Dx − xn , yn , Dx − xm , ym ) = τ0 (xn , yn , xm , ym ) τ0 (xn , Dy − yn , xm , Dy − ym ) = τ0 (xn , yn , xm , ym ) τ1 (Dx − xn , yn , Dx − xm , ym ) = τ1 (xn , yn , xm , ym ) τ1 (xn , Dy − yn , xm , Dy − ym ) = τ3 (xn , yn , xm , ym , Dy ) τ2 (xn , Dy − yn , xm , Dy − ym , Dy ) = τ2 (xn , yn , xm , ym , Dy ) τ2 (Dx − xn , yn , Dx − xm , ym , Dx ) = τ4 (xn , yn , xm , ym ) (7) Hence exchanging x by Dx − x or/and y by and Dy − y, with the appropriate mapping of the measured delays to the model, also leads to a valid result. Eventually, the estimation procedure may converge towards one of the 8 valid coordinate representations: (x, y), (y, x), (Dx − x, y), (y, Dx − x), (x, Dy−y), (Dx−x, Dy−y), (Dy−y, x) or (Dy −y, Dx −x). This shall not be viewed as a drawback however since we are only interested in relative positioning here (i.e. with respect to the room dimensions). B. Extended Kalman Filter A cooperative tracking filter involving 3 nodes and 2 links has been put forward. Accordingly, the state vector contains the 2D Cartesian coordinates and speeds of the 3 nodes, along with the room dimensions Dx and Dy . As for mobility, we consider Friedland’s model, which assumes constant velocities but random accelerations between consecutive measurements (see e.g. [9]). The state vector is then X(k/k) = [X1 (k/k) X2 (k/k) X3 (k/k) Dx Dy ] Xn (k/k) = [xn (kT ) x˙ n (kT ) yn (kT ) y˙ n (kT ) ], n = 1..3 (8)

The linear state transition matrix F from k − 1 to k is:   A 0 ... ... 0    . 1 T 0 0  0 A 0 0 ..    0 1 0 0     F =  ... 0 A 0 ...  with A =   0 0 1 T  (9)    .  0 0 0 1  .. 0 0 1 0  0 ... ... 0 1

where T is still the filter refreshment period. Following [9], the model error covariance matrix is:   G 0 ... 0 0  ..  0 G 0 0 .     .. ..  (10) Q= . 0 G 0 .     .   .. 0 0 σ 2 0  D 2 0 . . . . . . 0 σD  3  T /3 T 2 /2 0 0  T 2 /2 T 0 0  , σ 2 correwhere G = σa2   0 0 T 3 /3 T 2 /2  D 0 0 T 2 /2 T sponds to the model uncertainty for the room dimensions and σa2 characterizes the noise affecting the random acceleration model. The observation vector for the 3 cooperative links is: Z(k) = [z12 (k) z13 (k)]T

(11)

where z12 , z13 are the observation vectors constructed out of the measured delays {ti }i=0..4 over the 2 links involving node 1, rearranged with the mapping (ˆ a, ˆb) selected during initialization at k = 0 (See III-A). These vectors are related to the state vector variables for n = 1 and m = 2, 3 as follows: znm (k) = h(xm (k), ym (k), xn (k), yn (k), Dx , Dy ) + wnm (k) (12) In first approximation, the observation noise vector W (k) = [w12 (k)T w13 (k)T ]T is assumed to be zero-mean Gaussian with a covariance matrix R = σ02 I10 , where I10 is the 10x10 identity matrix and σ02 accounts for the quality of TOA estimation, and the measurement errors are assumed 2 independent. Advantageously σ02 should be related to σm . IV. S IMULATION R ESULTS A. Simulation Set-Up The proposed algorithm has been evaluated through simulations with the following assumptions. The true initial coordinates of 3 mobile devices are uniformly distributed on the intervals [0, Dx ] and [0, Dy ], respectively along the x and y axes. The nodes are moving with a true constant speed (i.e. assuming no reflections on the walls) whose components are uniformly distributed on [0, 0.5m/s] along each axis. The sampling frequency for channel estimation at the receiver is chosen as Fs = 250M Hz like in [8]. The EKF update period is T = 50ms. Considering synthetic measurements generated from Eq. (1) (though randomly unordered to account for mapping uncertainty), the standard deviation of the noise terms affecting the delay measurements in Eq. (12) spans from σm = 1/Fs = 4ns to σm = 2/Fs = 8ns, accounting for the internal time resolution available at the receiver (i.e. with an equivalent to 1 or 2-sample uncertainty on multipath TOA measurements). In the filter, the state model

2 error variances are set to σD = 10−4 and σa2 = 10−5 and the observation noise variance to σ02 = 1/Fs4 . Over each simulated mobile trajectory, the devices occupy 200 consecutive positions (i.e. simulating mobility over 10s). Finally, 200 distinct simulation runs are considered per scenario to draw empirical statistics. The performance is measured by means of empirical Cumulative Density Function (CDF) of the Root Mean Square Error (RMSE) of the estimated parameters (i.e. regarding Cartesian coordinates or Dx and Dy ), averaged over simulated trajectories. Since there are 8 valid coordinate representations that must be considered, due to valid flip/mirror and transmitter/receiver ambiguities (that we do not claim to solve anyway), 8 instantaneous error measurement vectors are computed at each time step k. The selected error vector is eventually defined by the configuration that leads to the smallest error. Finally, the minimization of the cost function Eq. (4) is performed with a classical gradient descent method [11].

B. Results On Fig. 2 we compare the CDF of the estimated variables RMSE with optimization-based and (genius-aided) perfect initialization procedures, for Vmax = 0.5m/s, Dx = 5m, Dy = 10m and different qualities of multipath delay measure2 ments (i.e. σm = 1/Fs2 vs. 4/Fs2 ). In spite of a purely blind initialization and reasonably conservative TOA precision levels (i.e. compatible with the time resolution usually available at real UWB receivers), the anchor-less location performance is on the order of that achieved by more conventional anchorsbased solutions, say with a median error better than 1m in all the tested cases. This would be compliant with most of indoor application requirements. Moreover, the proposed solution seems rather robust with respect to initialization, especially in terms of location errors. The degradation is however more significant for the estimated room dimensions, specially in the worst case error regime (i.e. when CDF(RMSE)=90%). Finally, Fig. 3 shows the impact of the room dimension Dx 2 = 1/Fs2 . (and dimensions ratio with respect to Dy ) for σm When the surface of the room is enlarged, the performance is improved due to the fact that the nominal real delays tend to increase while the received signal SN R (and hence, the multipath TOA measurement error) remains approximately constant in the room.

Fig. 2. Empirical CDF of RMS errors (in m) for estimated node locations (a) and estimated room dimensions (b) depending on the equivalent uncertainty on 2 = 1/F 2 vs. 4/F 2 ) and initialization (i.e. multipath TOA estimates (i.e. σm s s perfect vs. optimization-based search), with Vmax = 0.5m/s, Dx = 5m and Dy = 10m, for N = 3 nodes using L = 2 links.

V. C ONCLUSION AND P ERSPECTIVES We have presented a new method that enables to estimate the dimensions of a rectangular room and to perform simultaneously relative positioning of cooperative wireless devices out of received multipath profiles. The propagation delays corresponding to the direct and the single-bounce reflected paths are exploited within an original EKF formulation. On this occasion, we have mostly focused on purely blind initialization, data association (i.e. solving out the mapping ambiguity between the indexes of resolved multipath components and the a priori indexes of reflecting walls) and continuous tracking issues under mobility. Preliminary results obtained within synthetic simulation-based scenarios at relatively slow mobility (i.e. Vmax < 0.5m/s) show the benefits from cooperation (i.e. considering 3 nodes instead of simple single links like in [8]) and that rather satisfactory positioning and room dimensions estimation performances could be achieved for plausible variances of the estimated multipath delays (i.e.

Fig. 3. Empirical CDF of RMS errors (in m) for estimated node locations (a) and estimated room dimensions (b) depending on the actual room dimension along x (i.e. Dx = 5m or 10m), with Vmax = 0.5m/s, Dy = 10m and 2 = 1/F 2 , for the proposed optimization-based initialization and N = 3 σm s nodes using L = 2 links.

2 σm = 1/FS2 ). The performance of the proposed method also tends to improve with the dimensions of the room under constant TOA precision, as expected. As a first perspective to this work, considering that data association should be dissociated from the continuous multipath TOA single-link estimation problem, in a steady-state tracking regime (i.e. after initialization is performed), a popular Nearest Neighbor (NN) algorithm could be applied to update the mapping indexes (ˆ a, ˆb) at each time k, hence selecting the best observation candidate that is the nearest to the observation predicted by the tracking filter, i.e. following the reverse innovation monitoring approach (ˆ a, ˆb) = −1 T argmin(a,b) [V (k)S(k) V (k) ] where V (k) = [znm,ab (k) − zˆnm (k|k − 1)]T , zˆnm (k|k − 1) is the observation vector predicted according to Eq. (12) and S(k) is the filter innovation covariance matrix at time step k. Other axes of improvement should consider addressing the delays extraction problem with more realistic models regarding receiver capabilities, received multipath profiles and TOA measurement noise (i.e. instead of the synthetic measurements considered so far). Furthermore, the initialization step, which is still too demanding in terms of computational load (i.e. for implying numerous minimization tests) should be revised and alleviated. Finally, the proposed solution could be extended, by considering still L = 2 links but changing the roles at each of the N = 3 nodes (i.e. executing parallel calculations at nodes 1, 2 and 3) before exchanging and fusing local estimates afterwards (e.g. Dx , Dy ). The resulting impact on extra traffic and over-the-air signalling under mobility constraints shall however be carefully assessed.

ACKNOWLEDGMENTS This work has been carried out in the frame of the WHERE2 (FP7-ICT 248894) project, which is partly funded by the European Union. R EFERENCES [1] IEEE 802.15 Task Group 4, ”Part 15.4: Wireless medium access control (MAC) and physical layer (PHY) specifications for low-rate wireless personal area networks (WPANs)”, IEEE Std 802.15.4a-2007 document (Amendment to IEEE Std 802.15.4-2006), 2007. [2] A. Rabbachin, I. Oppermann, and B. Denis, ”ML Time-of-Arrival Estimation based on Low Complexity UWB Energy Detection”, in Proc. IEEE ICUWB06, pp. 599604, Waltham, Sept. 2006. [3] J. Youssef, B. Denis, C. Godin, S. Lesecq, ”New TOA Estimators within Energy-Based Receivers under Realistic UWB Channel Statistics”, in Proc. IEEE VTC-Spring’10, Taipei, May 2010. [4] P. Meissner, D. Arnitz, T. Gigl, K. Witrisal, ”Analysis of an Indoor UWB Channel for Multipath-Aided Localization”, in Proc. ICUWB’11, Bologna, Sept. 2011. [5] J. Seitz, et al., ”UWB Feature Localization for Imaging”, in Proc. IEEE ICUWB’08, 2008. [6] T. Deiler and J. Thielecke, ”Feature based indoor mapping using a bat type UWB Radar”, in Proc. IEEE ICUWB’09, pp. 475479, Vancouver, 2009. [7] W. Guo and N.P. Filer, ”On the accuracy of an indoor location-sensing technique suitable for impulse radio networks”, in Proc. IEEE ICC 07, Glasgow, June 2007. [8] V. La Tosa, B. Denis, B. Uguen, ”Joint Anchor-less Tracking and Room Dimensions Estimation through IR-UWB Peer-to-peer Communications”, in Proc. IEEE ICUWB’11, Bologna, Sept. 2011. [9] Y. Bar-Shalom and T.E. Fortmann, ”Tracking and Data Association”, Boston M.A: Academic, 1988. [10] K. Reif, S. Gnther, E. Yaz and R. Unbehauen, ”Stochastic Stability of the Discrete-Time Extended Kalman Filter”, IEEE Trans. On Automatic Control, Vol. 44, No. 4, April 1999. [11] S. Boyd and L. Vandenberghe, ”Convex optimization”, Cambridge University Press, 2004.

ICT–248894 WHERE2

A.13

D2.6

Cooperative Anchor-less Tracking and Room Mapping with Independent Computing Central Nodes and Application to Ultra Wideband Ray-Tracing Simulations

M. Des Noes and B. Denis.

120 / 146

Cooperative Anchor-less Tracking and Room Mapping with Independent Computing Central Nodes and Application to Ultra-Wideband Ray-Tracing Simulations M. des Noes, B. Denis CEA-Leti Minatec Campus 17 rue des Martyrs, 38054 Grenoble Cedex 09, France E-mails: [mathieu.desnoes, benoit.denis]@cea.fr

Abstract—In [4], a method was proposed to jointly estimate both the dimensions of a rectangular room and the relative locations of wireless devices communicating in this room. It exploits the 2 channel impulse responses between a central node and 2 other nodes. In this report, the method is extended to multiple central nodes by changing the roles at each of the N = 3 nodes. Then, the algorithm is evaluated with synthetic channel profiles generated through UWB Ray-Tracing simulations according to the building layout in the reference synthetic T2.3 environment (SIR premises). Index Terms—Extended Kalman Filter, Impulse Radio, Localization, Mapping, Wireless Sensor Networks, Ultra Wideband.

I. I NTRODUCTION Many emerging applications such as context-aware network optimization, self-reconfigurable Hi-Fi systems, augmented reality in gaming or assisted navigation in confined environments, require that mobile wireless nodes perform simultaneously self-localization and environment mapping while communicating with limited access to elements of infrastructure. In [4], a new method was proposed to jointly and simultaneously estimate both the dimensions of a rectangular room (i.e. mapping) and the relative locations of communicating devices in this room (i.e. relative positioning), coping with various pairs of mobile devices in a coordinated and cooperative context. The method takes benefit from the multipath propagation delays classically observed over radio links in indoor environments. The processing of these delays, which naturally convey information about the geometry of the involved radio links, makes possible the estimation of the room dimensions and the coordinates of the communicating nodes in this room. In this contribution, the algorithm proposed in [4] is extended by considering still L = 2 links but changing the roles at each of the N = 3 nodes (i.e. executing parallel calculations at nodes 1, 2 and 3) before exchanging and fusing local estimates afterwards (e.g. Dx , Dy ). Then the previous version of the algorithm (without changing the role of each node) is evaluated based on the channel impulse responses generated by UR1 given a realistic environment. These synthetic channel profiles were generated through UWB Ray-Tracing simulations according to the building layout in the reference synthetic T2.3 environment (SIR premises). The document is organized as follows. Section II describes the system model as an input to our algorithm definition.

Section III details the extension of the algorithm to multiple central nodes. Section IV illustrates the performance of the proposed solution in realistic scenarios, and Section V concludes the report. II. S YSTEM M ODEL As shown on Figure 1, we first assume two nodes communicating in a rectangular room over Line Of Sight (LOS) radio links, including a direct path and 4 single-bounce reflections on surrounding walls. Standard 2D Cartesian coordinates (xn , yn ) and (xm , ym ) are used to represent the locations of nodes n and m, respectively. The constant dimensions of the room are noted Dx and Dy along the x and y axes respectively. At each discrete time instant kT (with T the refreshment period), the theoretical delays associated with the 5 multipath k components {τmn,j }j=0..4 mentioned above are related to the 6 unknown geometrical variables as follows:  k T k k k k (τmn,0 )2 (τmn,1 )2 (τmn,2 )2 (τmn,3 )2 (τmn,4 )2 = 

1 c2

(xm (k) − xn (k))2 + (ym (k) − yn (k))2



    2 2   (xm (k) − xn (k)) + (ym (k) + yn (k))        (2Dx − xm (k) − xn (k))2 + (ym (k) − yn (k))2         (xm (k) − xn (k))2 + (2Dy − ym (k) + yn (k))2      2 2 (xm (k) + xn (k)) + (ym (k) − yn (k))

(1) where c = 3.108 m/s is the celerity of light, n and m are the indexes of the 2 communicating nodes and where the reflecting surfaces are numbered according to an arbitrary order j = 0..4. In real-life systems, the multipath delay profile between each pair of devices may be estimated using a receiver compatible with the IEEE 802.15.4a standard [1]. For instance the delays associated with some of the strongest energy bins in a non-coherent energy detector might be considered [2]. As the multipath TOA procedure does not fall in the scope of this very first part of our investigation, we simply assume that 5 time measurements affected by Additive White Gaussian Noise 2 (AWGN) terms {bkmn }, which are i.i.d. with a variance σm for all the measurements, are squared into unordered observations {tkmn } as follows: k tkmn,i ←→ (τmn,j + bkmn,j )2 , i, j = 0..4

(2)

It is hence worth noting that the squared delays measured at the receiver are indexed by a variable i = 0..4, which may not coincide with the index j = 0..4 of the prior wall numbering model in Eq. (1). This is due to the fact that the receiver has no a priori knowledge of the correct mapping association.

of the Root Mean Square Error (RMSE) of the estimated parameters (i.e. regarding Cartesian coordinates or Dx and Dy ), averaged over simulated trajectories. Since there are 8 valid coordinate representations that must be considered, due to valid flip/mirror and transmitter/receiver ambiguities (that we do not claim to solve anyway), 8 instantaneous error measurement vectors are computed at each time step k. The selected error vector is eventually defined by the configuration that leads to the smallest error. Finally, the minimization of the cost function (see [4]) is performed with a classical gradient descent method [3]. B. Results with synthetic multipath

Fig. 1. Single-bounce reflection model for 2 nodes communicating over Line of Sight links in a rectangular room, with an arbitrary wall numbering (*).

Fig. 3 and Fig. 4 shows the impact of the multi-central nodes procedure for Vmax = 0.5m/s. There is no observed benefit for the location and the room dimensions estimation. This is due to the application of this concept only to the state initialization step of the algorithm. The estimation gain at initialization is small and the overall performances are mainly driven by the noise over the nodes trajectory.

III. E XTENSION TO MULTIPLE CENTRAL NODES In [4], a node was selected as being “central” and the 2 communication links with the 2 other nodes were exploited to estimate simultaneously the room dimensions and the nodes locations. This concept has been generalized and is depicted in Fig. 2. Each node is successively considered as been central, and the state initialization procedure described in [4] is applied. The configuration which gives the minimum cost function is selected, and the corresponding initial state and configuration are used by the tracking filter. Fig. 3. Empirical CDF of RMS errors (in m) for estimated node locations with the multi-central nodes procedure, with Vmax = 0.5m/s, Dx = 5m, 2 = 1/F 2 Dy = 10m and σm s

Fig. 2.

Multiple central nodes.

IV. S IMULATION R ESULTS A. Simulation Set-Up The proposed algorithm has been evaluated through simulations with the following assumptions. The true initial coordinates of 3 mobile devices are uniformly distributed on the intervals [0, Dx ] and [0, Dy ], respectively along the x and y axes. The nodes are moving with a true constant speed (i.e. assuming no reflections on the walls) whose components are uniformly distributed on [0, 0.5m/s] along each axis. The sampling frequency for channel estimation at the receiver is chosen as Fs = 250M Hz. The EKF update period is T = 50ms. 2 The state model error variances are set to σD = 10−4 and 2 −5 σa = 10 and the observation noise variance to σ02 = 1/Fs4 . Over each simulated mobile trajectory, the devices occupy 200 consecutive positions (i.e. simulating mobility over 10s). Finally, 200 distinct simulation runs are considered per scenario to draw empirical statistics. The performance is measured by means of empirical Cumulative Density Function (CDF)

Fig. 4. Empirical CDF of RMS errors (in m) for estimated room dimensions with the multi-central nodes procedure, with Vmax = 0.5m/s, Dx = 5m, 2 = 1/F 2 Dy = 10m and σm s

C. Results with a realistic environment The proposed algorithm has also been evaluated with more realistic received channel profiles obtained by means of UWB Ray-Tracing simulations (T2.3 Ray-Tracing simulations provided by IETR/INSA). The measurement environment is shown in Fig. 5. The transmit locations are marked with a red point, while receiver locations are in blue. The algorithm has been evaluated based on the channel impulse responses measured in the room delimited by the rectangle colored with red. At first, the channel impulse responses were integrated on a 1 ns window, then the 10th strongest multipath were

selected as potential candidate. It was shown that 40% of the channel impulse responses could not be exploited to solve our problem. This is shown in Fig. 6. The locations where the measured CIR is not exploitable are marked with a red circle, while those in green are exploitable. In the non exploitable configurations, at least one multipath corresponding to an expected reflection is not present. An exemple of such a configuration is given in 7. The last theoretical multipath is not selected among 10th strongest multipath. Among the 60% of exploitable channel impulses responses, some configurations have been tested for a practical evaluation of the algorithm. The algorithm needs to select 4 out of the 9th strongest in order to feed the algorithm with the estimated delays corresponding to the single reflections on the walls. It is assumed that the first multipath corresponds to the direct transmission. As a consequence, a total of C94 = 126 multipath configurations must be tested. The one giving the smallest initial cost function is selected. In this document, the results obtained with 2 configurations are reported: • Configuration 1 corresponds to the transmitter (Tx) and receiver locations 12 and 28 in Fig. 6. • Configuration 2 corresponds to the transmitter (Tx) and receiver locations 13 and 28. 1) Configuration 1: The selected multipath configuration gives a cost function error equals to 0.24 which is quite large. Tab. I and Tab. II show the estimated room dimensions and node locations. The estimation error is very large. Dx - real 8.5

Dx - estimated 4.1

Dy - real 4.4

Dy - estimated 5

TABLE I ROOM DIMENSION ESTIMATIONS - CONFIGURATION 1

Fig. 5.

Fig. 6.

Measurement environment

Measurement grid in the selected room

3) Conclusion: The results obtained with 2 configurations shows that: • •

The estimation error is very marge. The simple delay extraction method used in these simulation is not adequate. Even if the initial state estimation gives a small cost function error (e.g. configuration 2), the algorithm may give a large estimation error. In fact, the selected multipath configuration (1 out of the 126 possibilities) may correspond to a valid by different system configuration (room dimensions and node locations). V. C ONCLUSION AND P ERSPECTIVES

(x,y) - real (x,y) - estimated

Tx (4 , 3.5) (-0.4 , 2.3)

Rx12) (3.1 , 0.5) (-2.6 , 0.2)

Rx28 (6 , 2.5) (-1.4 , 4.2)

TABLE II L OCATION ESTIMATIONS - CONFIGURATION 1

2) Configuration 2: The selected multipath configuration gives a cost function error equals to 0.0042 which is very small. Tab. III and Tab. IV show the estimated room dimensions and node locations. As for the configuration1, the estimation error is very large. Dx - real 8.5

Dx - estimated 6.1

Dy - real 4.4

Dy - estimated 4.9

TABLE III ROOM DIMENSION ESTIMATIONS - CONFIGURATION 2

(x,y) - real (x,y) - estimated

Tx (4 , 3.5) (0.44 , 2.4)

Rx13) (3.1 , 1.5) (0.8 , 0.3)

Rx28 (6 , 2.5) (1.6 , 4.1)

TABLE IV L OCATION ESTIMATIONS - CONFIGURATION 2

In [4], we have presented a new method that enables to estimate the dimensions of a rectangular room and to perform simultaneously relative positioning of cooperative wireless devices out of received multipath profiles. The propagation delays corresponding to the direct and the single-bounce reflected paths are exploited within an original EKF formulation. The proposed method has been extended by considering still L = 2 links but changing the roles at each of the N = 3 nodes (i.e. executing parallel calculations at nodes 1, 2 and 3) before exchanging and fusing local estimates afterwards (e.g. Dx , Dy ). Simulation results show that this cooperation technique is inefficient if it is only applied to the initial state estimation. The proposed algorithm has also been evaluated with more realistic channel profiles obtained by means of UWB RayTracing simulations. Preliminary simulation results exhibit poor estimation performance. This is partly due to the delays extraction procedure which was not adequate. In addition, the selected multipath configuration (1 out of the 126 possibilities) may correspond to a valid by different system configuration (room dimensions and node locations). As a first perspective to this work, the initial state estimation problem should be revisited. This must be done jointly with the extraction of multipath delays.

Fig. 7.

Not exploitable CIR - Rx location N 6

ACKNOWLEDGMENTS This work has been carried out in the frame of the WHERE2 (FP7-ICT 248894) project, which is partly funded by the European Union. R EFERENCES [1] IEEE 802.15 Task Group 4, ”Part 15.4: Wireless medium access control (MAC) and physical layer (PHY) specifications for low-rate wireless personal area networks (WPANs)”, IEEE Std 802.15.4a-2007 document (Amendment to IEEE Std 802.15.4-2006), 2007. [2] V. La Tosa, B. Denis, B. Uguen, ”Joint Anchor-less Tracking and Room Dimensions Estimation through IR-UWB Peer-to-peer Communications”, in Proc. IEEE ICUWB’11, Bologna, Sept. 2011. [3] S. Boyd and L. Vandenberghe, ”Convex optimization”, Cambridge University Press, 2004. [4] M. des Noes and B. Denis, ”Benefits from Cooperation in Simultaneous Anchor-less Tracking and Room Mapping based on Impulse Radio - Ultra Wideband Devices”, in Proc. IWSSIP, Vienna, April 2012.

ICT–248894 WHERE2

A.14

D2.6

Extraction of Context-Aware Features and Localization from Indoor-to-Outdoor Data

Julien Stéphan, Yoann Corre Extraction of context-aware features and localization from indoor-to-outdoor data. Internal Report., April , 2013, Rennes, France.

125 / 146

WHERE2

Extraction of context-aware features and localization from indoor-to-outdoor data Julien Stéphan, Yoann Corre Radio Business Unit SIRADEL, Rennes, France Investigations on the construction of environment maps are usually limited either to the indoor or outdoor environment. However using this kind of technique to geo-locate the indoor terminals in a global environment (in which building? where inside the building?) might also be of a great interest. That is why we investigate an innovative technique that consists in the extraction of context–aware features from indoor-to-outdoor channel data. The main idea lies in finely analyzing radio wave propagation when one terminal is inside a building and another terminal is moving outside in close vicinity of the building and extracting coarse information about the building (e.g. building shape, presence of windows). We study in WHERE 2 the feasibility of such a technique and search after the channel parameters that may provide the most relevant information. At this stage, investigations are based only on channel simulations. The technique is more precisely based on the detection of the static clusters that may for instance be generated by the window/door frames. The investigations make use of the original indoor-outdoor synthetic test environment presented in D2.3 [1], the highly realistic indoor-outdoor channel model elaborated in WP1 (see [2] for more details) and an algorithm to detect the presence of static clusters in successive (in space or time domain) channel responses described in D1.8 [2]. This cluster detection algorithm gather multiple simulated ray paths based on angular and delay spreads criteria. Basically, we consider a time-variant channel related to a succession of radio links composed of a static indoor terminal and a moving outdoor terminal. We use a two-step process where we first extract the main clusters on local intervals (5 meters wide in the current study). Then, we track the clusters over several intervals to identify those that are persistent. Similar clusters in two successive intervals are identified from the correlation between their delay, angular and power statistics. The analysis has been done from three different Access Points (as indoor terminal) set in the same room in the first floor of the main building (see Figure 1). The distance between the building façade and access points is 60 cm, 2.10 m and 3.85 m for respectively AP1, AP2 and AP3. The paths perpendicular to the façade passing respectively by AP1 and AP3 cross the façade at a concrete wall. The path perpendicular to the façade passing by AP2 crosses the façade at a window. We consider a moving terminal outdoors, which follows a 90 m-long linear trajectory in front of the main building (blue arrow in the figure). An omnidirectional antenna is considered for all APs as well as for the moving terminal. It is set at the height of 1.85 m for both (remark that windows are present at 1.0 m above the ground floor). Simulations have been computed at the frequency of 2462 MHz (i.e. channel 11 of the WiFi standard).

Building Vegetation Concrete Plasterboard Glass Column (concrete) Wood

25m

AP 1 Concrete Plasterboard Glass Column (concrete) Wood

AP 3 AP 2

7m

AP 1

3.85m

AP 2

AP 3 1.75m

Figure 1: Outdoor-indoor radio link scenario.

Persistent clusters have been extracted from the successive radio links simulated between each Access Point (AP) and the moving outdoor terminal. Figure 2 shows the power, mean angle-of-departure, mean angle-ofarrival and mean delay of these clusters varying along the trajectory of the moving user terminal. 6 persistent clusters have been extracted for AP 1 and 8 for AP 2 and AP 3. The cluster related to the LOS direct paths is also represented by the red dotted line. This cluster is very dominant when existing, making all other main clusters almost negligible. Its existence is limited to only 2 intervals (thus approximately 10 m) for AP1 and AP3. It is longer for AP2, i.e. 6 intervals (approximately 30 m). The difference is due to the position of the indoor terminal relative to the window. We imagine this specific behavior may be used as a fingerprint to extract information on the location of the AP relative to close apertures. (a) AP1

Cluster ID 1 2 3 4 5 6 DP

(b) AP2

Cluster ID 1 2 3 4 5 6 7 8 DP

(c) AP3

Cluster ID 1 2 3 4 5 6 7 8 DP

Figure 2: Evolution of cluster properties along the moving terminal trajectory.

Some of persistent clusters have a constant angle properties (angle-of-departure or in angle-of-arrival). We finely characterized those that are present over at least 5 successive intervals, i.e. the clusters with ID 1, 2 and 3 for AP1 and AP2 and the clusters with ID 1, 3 and 8 for the AP3. We calculate the location of the interaction point or “bright spot” related to the angle-of-departure and angle-of-arrival mean properties of each cluster for each interval (i.e. location of the outdoor mobile terminal every 5m along the trajectory), considering that the main contribution(s) in high-power clusters are more likely to result from one single interaction, and identify with which physical element of the geo map data (window frames notably) it can be associated. Remark that, in our approach, we know the exact trajectory of multipath that composed each cluster as they are computed by a raybased model. We thus can confirm that multipath that compose these clusters have an unique interaction with the environment (e.g. a reflection or diffraction). In a scenario with real measurements, we would have to make that assumption most likely based on the relative high power of detected clusters. Figure 3 illustrates the mean location of bright spots obtained for each AP. Correlation with specific elements of the map data, either a window or a building edge, is obvious. Table 1 details the distance mean error between bright spots of each cluster and exact location of the related element of the map data. The global mean error is 0.97 m only, and standard deviation is 1.17 m. The mean error grows up to 9.59 m if we consider only the angle-of-arrival statistics of each cluster over the successive intervals. Table 1: Mean error of location of the geo map data elements.

AP1 Mean Error Standard deviation error

AP2

AP3

Cluster 1

Cluster 2

Cluster 3

Cluster 1

Cluster 2

Cluster 3

Cluster 1

Cluster 3

Cluster 8

0.64 m

0.85 m

3.02 m

0.68 m

0.07 m

0.81 m

0.72 m

0.67 m

1.23 m

0.89 m

2.01 m

1.14 m

0.40 m

0.06 m

0.68 m

0.34 m

0.57 m

0.22 m

(a) AP1

(b) AP2

(c) AP3

Figure 3: Bright spots for Access Point 1, 2 and 3.

Lastly, we consider the scenario where the AP locations are unknown. We thus try to estimate it by assuming exact knowledge of the moving user terminal location (i.e. a perfect GPS location outdoors) and using the mean delay statistic, mean angle-of-departure statistic and mean bright spot location obtained for each extracted static cluster. Table 2 details the distance error obtained for each AP by considering each static cluster individually. The obtained estimation is good: the mean error on the AP location is globally 1.47 m. This result confirms the feasibility and interest of the proposed technique. Moreover, we expect that utilization of several static clusters conjointly would further improve the estimation. Table 2: Mean error of AP location estimation.

Cluster 1 0.76 m

AP1 Cluster 2 0.51 m

Cluster 3 4.89 m

Cluster 1 1.62 m

AP2 Cluster 2 0.07 m

Cluster 3 2.57 m

Cluster 1 2.49 m

AP3 Cluster 3 0.05 m

Cluster 8 0.35 m

Another approach would be to use conjointly the knowledge of the geo map data and only a limited set of cluster properties (e.g. power and delay, or power and angle-of-arrival). In frame of the work package 4, we propose a novel technique based on the utilization of an advanced GPS, three 2D lasers and five cameras set on a specific mobile mapping vehicle to automatically create or refine an outdoor digital geo map data from field measurements. This technique will be fully detailed in the last WP4 deliverable. A possible perspective for this work will be thus to combine both proposed techniques to derive an even more precise SLAM algorithm for detection of indoor terminal locations based on indoor-outdoor measurements. Possible application may be to combine such a technique with wide band and Multiple Input Multiple Output antenna based outdoor terminals (typically LTE/LTE-A mobile phone or mobile sounders) equipped with GPS and able to scan outdoor/indoor networks (e.g. 4G plus WiFi) in order to localize the indoor APs in a global environment. Then the localized indoor APs can help in the global localization of the indoor terminals attached on them.

REFERENCES [1] [2] [3]

WHERE2 Partners, “Deliverable D2.3: Intermediate report on Self-learning positioning using inferred context informationScenarios and parameters”. Deliverable FP7-ICT248894, WHERE2, 2011. WHERE2 Partners, “Deliverable D1.6: Ray-tracing tool for heterogeneous distributed positioning”. Deliverable FP7-ICT248894, WHERE2, 2012. WHERE2 Partners, “Deliverable D1.8: Final report on the WHERE2 Channel Model”. Deliverable FP7-ICT248894, WHERE2, 2012.

ICT–248894 WHERE2

A.15

D2.6

Hidden Markov Model based Mobility Learning for Improving Indoor Tracking of Mobile Users

T.Laursen, N. B. Pedersen, J. N. Nielsen, and T. K. Madsen In Proceedings of The 9th Workshop on Positioning, Navigation and Communication 2012 , Special Session on Localization and Tracking, Dresden, Germany, March 2012. c

2012 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

134 / 146

Hidden Markov Model based Mobility Learning for Improving Indoor Tracking of Mobile Users Troels Laursen, Nikolaj Bisgaard Pedersen, Jimmy Jessen Nielsen, and Tatiana K. Madsen [email protected], [email protected], [email protected], [email protected] Section for Networking and Security, Department of Electronic Systems, Aalborg University, Denmark

Abstract—Indoors, a user’s movements are typically confined by walls, corridors, and doorways, and further he is typically repeating the same movements such as walking between certain points in the building. Conventional indoor localization systems do usually not take these properties of the user’s movements into account. In this paper we propose a Hidden Markov Model (HMM) based tracking algorithm, which takes a user’s previous movements into account. In a quantized grid representation of an indoor scenario, past movement information is used to update the HMM transition probabilities. The user’s most likely trajectory is then calculated using and extended version of the Viterbi algorithm. The results show significant improvements of the proposed algorithm compared to a simpler moving average smoothing.

I. I NTRODUCTION Indoor localization of mobile user’s needs to rely on radiolocation methods, since satellite-based navigation systems such as GPS usually do not have the required view to the sky to properly localize. Even with local indoor radiolocation systems that use RSS or ToA/TDoA measurements of available wireless networks, e.g., a WiFi network based on IEEE 802.11n which recently had an amendment to support TDoA measurements for ranging, will experience some level of localization errors due to the impact of walls, furniture and other obstructions. Typically a localization system is composed of the components shown in Fig. 1.

Fig. 1.

Localization system overview

Initially, the radio measurements are fused or otherwise processed to be converted to range estimates and/or converted into geographical coordinate estimates, using geometrical techniques such as trilateration, triangulation, etc. However, noisy measurements necessitate further filtering of the location estimates. Common approaches for this are the Kalman and particle filters, as described in [1], [2]. Indoors, a user is often repeating the same movements, such as walking between main door, office, rest rooms, canteen, printer, and colleagues’ offices. State of the art localization approaches such as the Kalman and particle filters do not This work has been performed in the framework of the ICT project ICT248894 WHERE2, which is partly funded by the European Union.

take this determinism of the user into account. The main aim of the work described in this paper is to propose a method for accounting for a user’s repeated behavior in the location determination process. In the literature there are several examples of existing work where learning movement patterns is exploited. Some examples are listed in the following: BreadCrumbs [3] tracks the movement of the device’s owner using GPS, and customizes a predictive mobility model for that specific user. The learned mobility model is a 2nd order Markov Chain, where each state represents a 2D geographical grid location covering 110m ⇥ 80m. Additionally, also the previous location is a part of the state information. The purpose of the protocol is to predict connectivity options a number of steps into the future. This is done by weighting the achievable bandwidth in possible future states with the probability of being in those states, which is given by the Markov Chain transition probabilities. The authors in [5] compare different methods for prediction a user’s next location in an indoor office environment. Specifically, they compare dynamic Bayesian network, multilayer perceptron, Elman net, Markov predictor, and state predictor. For evaluation they use the Augsburg Indoor Location Tracking Benchmarks. This work did however not specifically consider the movement trajectory of the user, only what his next location would be and when he would most likely enter this location. Markov Models of different orders, i.e., Markov models where not only the current, but a number of previous states are taken into account, are in [7] considered for movement prediction. The authors test the proposed algorithms on two sets of measurements, namely Augsburg Indoor Location Tracking Benchmarks and Nokia Context Data, which contain the movements of different users over several months. The models are shown to work well in predicting the sequence of states that defines a user’s movements. Notice however that the user’s speed is not modeled. In [8] the authors propose a localization algorithm, which uses a short history of past RSSI measurements of a moving user to estimate the moving user’s location. Their approach requires that the system is first trained using movement paths with ground truth knowledge. In the testing phase, the authors have used different paths, composed of segments of the testing paths. They use radial basis function (RBF) fitting to learn a

reliable estimate of a mobile node’s position given its past signal strength measurements. They show that using historical data improves the localization accuracy by almost a factor of two compared to using only the most current measurements, in an indoor setting where MicaZ motes are used. An extended version of the fingerprinting based system RADAR proposed in [10] is proposed in [9]. Specifically, the authors have added information about the possible movements of a user. For example, the user would not be able to walk through a wall. This added environment knowledge is shown to slightly improve localization performance. The information about walls is however not learnt, but provided a priori. Nevertheless, this work indicates the possible improvement. In this paper we propose a method for mobility learning and subsequent filtering of location information improving tracking of mobile users indoors that is based on the Hidden Markov Model (HMM) framework, see [11]. First, in section II we describe the considered radio-location based system. Hereafter in section III we introduce the concepts and notation of the HMM framework, as well as the proposed extensions that enable mobility learning. For assessment of the proposed extensions, we present a test scenario in section IV, for which the test results are given in section V. Finally, the conclusions and outlooks are given sin section VI. II. S YSTEM D ESCRIPTION Assume that Z is a 2D grid representation of a geographical area, where each grid point corresponds to a geographical grid square. Within this area there is a mobile user, whose location is being tracked by a radio-based localization system which gives the location of the user at time t as a grid coordinate (x(t), y(t)). That is, we assume that the user’s estimated location is quantized according to the grid resolution and whenever the user is located inside a grid square, his/her location corresponds to that grid point. Clearly, the grid resolution sets a limit on the achievable accuracy of the localization system, which should be taken into consideration when evaluating the performance of the system. In practice, the user’s location estimate is already inaccurate due to for example propagation effects caused by nearby objects and obstructions if the localization system is for example signal strength or time-of-flight based. The inaccuracy due to the grid discretization may therefore not be significant compared to the error of the localization system. Over time, the user’s movement history can be represented as a sequence of location estimates provided by the localization system. Radio measurements, e.g., RSS, ToA

Localization system

(x(t),y(t))

HMM model

(x*(t),y*(t))

Fig. 2. The proposed HMM based model takes the position estimates (x(t), y(t) from an existing localization system as input and uses mobility learning to deliver an enhanced output (x ⇤ (t), y ⇤ (t).

The block diagram in Fig. 2 shows how the HMM based model works as a mobility learning extension to an existing

localization system. In the following section we describe how the HMM framework is applied to achieve mobility learning. III. H IDDEN M ARKOV M ODEL FOR M OBILITY L EARNING A Hidden Markov Model (HMM) is characterized by the following [11]: • S = {S1 , S2 , S3 , . . . , SN } is the set of possible states, where a state corresponds to a geographical grid point in Z. The principle is sketched in Fig. 3. • A = {ai,j } is the state transition probability distribution between states i and j, where ai,j = P [qt+1 = Sj |qt = Si ], 1  i, j  N and qt is the actual state at time t. Since we allow only movements between neighboring states, only the neighbor and self transition probabilities are non-zero. • ⇡ = {⇡i } is the initial state probability distribution, where ⇡i = P [q1 = Si ]. • V = {v1 , v2 , v3 , . . . , vM } is the set of possible observation symbols, which in our case corresponds to the set of grid locations. In this work we have the special situation that our symbols correspond to the possible states. • O = {O1 O2 O3 . . . OT } is a sequence of observations, where each Ot takes the value of a possible state in S. In our case O is a sequence of measured geographical grid locations made available from the localization system. • B = {bj (k)} is the observation symbol probability distribution in state j, where bj (k) = P [vk at t|qt = Sj ], 1  j  N, 1  k  M.

Fig. 3.

Markov chain representation of geographical area.

A. Viterbi Algorithm Initially we consider simply the problem of determining the most likely path Q = {q1 q2 q3 . . . qT }, defined as a sequence of states from S, given a sequence of observations O = {O1 O2 O3 . . . OT }. This problem can be solved using the Viterbi algorithm [11], which has the following steps: 1) Initialization: 1 (i) 1 (i)

= ⇡i bi (O1 ), 1  i  N = 0.

(1) (2)

2) Recursion: t (j)

= max [ 1jN

1 t  T

t 1 (i) ai,j ] bj (Ot ),

1 i  N (3) t (j)

= arg max[ 1jN

2 t  T

t 1 (i) ai,j ],

1 i  N. (4)

3) Termination: qT⇤

= arg max[ 1iN

which is a simple function that maps the past movement into a direction represented by the numbers in Fig. 4. For the directional Viterbi algorithm we have considered the following three different ways of constructing the transition probability matrix A. 1) All directions are equally likely: For transitions from a state i, ai is all-zero, except for the transitions to the same state or to the neighboring grid points. The matrix in eq. (10) corresponds to movement directions in Fig. 4.

(5)

T (i)].

ai,j

4) Path (state sequence) backtracking: qt⇤ =

⇤ t+1 (qt+1 ),

t=T

1, T

2, · · · , 1,

(6)

where ⇡i is the initial state probability for state i; t (i) is the best score (highest probability) along a single path, at time t, which accounts for the first t observations and ends in state Si ; t (j) is an array, which is used to keep track of the argument, i.e., the value of i, that maximized eq. (3); and qt⇤ is the state at time t for the best state sequence.

2

3 0.0625 0.0625 0.0625 0.5 0.06255 = 40.0625 0.0625 0.0625 0.0625

2) Linear movements: For transitions from a state i, ai is all-zero, except for the transitions to the same state or to some of the neighboring grid points. The matrix in eq. (11) corresponds to movement directions in Fig. 4, and the matrix is rotated according to the direction of movement in the past transition. If the past transition was to stay in the same geographical grid point, the movement direction from the transition before that is used.

B. Directional Viterbi Algorithm In this paper we propose to take the user’s habits into account by extending the state space to take the user’s previous movement direction into account. Specifically, for the proposed directional Viterbi algorithm, we have used a slightly modified version of the Viterbi algorithm where instead of ai,j we use ai,j,d . Here d = {1, 2, 3, 4, 5, 6, 7, 8, 9} corresponds to the movement direction in the past transition (t 1 ! t) as defined in Figure 4. This dependency on the previous movement direction means that the size of the state space is increased from N states to N ⇥ 9 states.

1

2

3

4

5

6

7

8

9

= max [ 1jN

t 1 (i) ai,j,d ] bj (Ot ),

1 t  T 1 i  N

t (j)

= arg max[ 1jN

t 1 (i) ai,j,d ],

where d is given by a function f : d = f(

t 1 (i), j),

(7)

2 t  T 1 i  N,

2

0.05 = 4 0.0 0.0

0.125 0.5 0.0

3 0.15 0.1255 0.05

(11)

3) Update from user movement traces: Here A is dynamic and updated according to either training data or continually from past user movements. Given a set of training data in the form of a sequence of states Q = {q1 q2 q3 . . . qT }, and a sequence of direction changes D = {d2 d3 d4 . . . dT } computed from the transitions in Q, we update the transition probability matrix A as follows: 1 ,qt ,dt

(t) = aqt

1 ,qt ,dt

(t

1) · (1 + wf )

(12)

where wf is the forgetting factor, which is used to control how much newer measurements are weighted versus old measurements. For each probability update we need to re-normalize the probabilities in the ”direction subspace” to sum to 1: a qt

1 ,qt

(t) =

a qt

(t 1) 1 ,qt 1 + wf

(13)

IV. T EST S CENARIO

The recursion step in eq. (3) and (4) is modified as: t (j)

ai,j

aq t

Fig. 4. Extended state space to indicate direction of movement, namely either to stay in the current grid cell or to move to one of the neighboring grid cells.

(10)

(8)

(9)

For evaluation of the proposed algorithms, we have used simulations of user movements between the way points illustrated in the building layout in Fig. 5. As shown in the figure, the considered building is represented in a quantized grid, with dimensions of 50 ⇥ 50 square grid units, where each grid cell corresponds to a state in S. The total number of states is 50 · 50 = 2500. The choice of representing the building map by a 50 ⇥ 50 cells grid is decided upon as a trade off between the following effects: 1) the finer the grid, the better the resolution of the location estimates; 2) the higher the number of states, the more time-consuming solving the

HMM using the Viterbi algorithm becomes; 3) in order to reflect the user’s past behavior with the state space extension depicted in Fig. 4, the grid size and sample time should be chosen jointly so that the user’s typical movement between two time samples is translated into a movement from one grid cell to a neighboring grid cell. The set of possible observation symbols V corresponds to the grid cells in the map.

In the evaluation, we assume that the localization system provides position estimates subject to 2D zero-mean Gaussian 2 . This is illustrated in the block noise with variance pos diagram in Fig. 7. In this study, it has been assumed that the 2 2 noise level is known, i.e., pos = sym .

The map with all the way points

Fig. 7. Block diagram for the simulation model. Y (n) is the non-observable true location, X(n) is the observable noisy location, after the 2D Gaussian noise factor W (n) has been added.

5

Grid numbers in Y−direction

10 15 20

V. R ESULTS AND D ISCUSSION

25 30 35 40 45 50 5

Fig. 5.

10

15

20 25 30 35 Grid numbers in X−direction

40

45

50

Building map, with walls and way points outlined.

The user movement trajectories are generated between the way points in Fig. 5 using the shortest path algorithm A*, which was first published in [12] and is widely used in graph theory for finding the shortest path between two nodes in a graph. In total 34 different trajectories have been created. A subset of the resulting movement trajectories that share the same starting way point are shown in Fig. 6. Notice how the A* star algorithm guides the movement trajectories through door openings towards the destination way point and not through walls. Map with the all the paths starting from one way point

5

Grid numbers in Y−direction

10 15 20 25

In this section the proposed HMM/Viterbi based algorithms are evaluated. First they are evaluated for a set of fundamental movement patterns and second they are evaluated in the test scenario outlined in the previous section. For reference, the algorithms are compared to a simple moving average filter (4th order). The considered performance metric is the average error, where the error is calculated as the difference between the true position and the estimated position, shown in Fig. 7 as Y (n) and X(n), respectively. A. Fundamental Movement Patterns Initially, the proposed algorithms have been evaluated for a set of fundamental movements, namely: straight line, 90 turn, 180 turn, and circle, as depicted in Fig. 8(a). The corresponding results are summarized in Fig. 8(b). It is immediately clear from the results that all of the Viterbibased algorithms outperform the simple moving average filter. Further we see that the learning Viterbi algorithm is performing better than the two other Viterbi algorithms, except in the case of straight line movements. However, in this case and in the case of circular turns the difference between the algorithms is almost negligable, whereas for the remaining two cases of 90 and 180 degree turns, the difference is more pronounced. Having seen some potential improvements for the Viterbi based algorithms for these fundemental but maybe also artificial movements, we consider in the following some more realistic simulated pedestrian movements.

30

B. Pedestrian Movement Traces

35

In the following, the results of the considered algorithms are shown for online and offline modes. In online operation, which corresponds to a localization or tracking application, the estimated trajectory is adapted progressively over time as more measurements become available, and here the error the error is calculated as the average over the considered time window. As opposed to that, when performing post-processing in offline mode, the algorithms have all measurements available for determining the most likely trajectory. The results shown in Table I compares the results of the different approaches for online and offline operation, as mean values of 10 noise realizations of each of the 34 possible movement trajectories.

40 45 50 5

10

15

20 25 30 35 Grid numbers in X−direction

40

45

50

Fig. 6. Movement trajectories originating from a single way point, before the artificial location error is added.

The observation symbol probability distribution B is defined 2 as a 2D zero-mean Gaussian distribution with variance sym around the geographical grid points that correspond to each observation symbol.

VI. C ONCLUSIONS AND O UTLOOK

(a) Movements 1 Moving Average Viterbi − any direction Viterbi − prefer linear Viterbi − learning

0.9

Avg. error [grid points]

0.8 0.7

We have presented a hidden Markov model based approach for improving accuracy of a moving object tracking in an indoor environment. This approach is completely decoupled from a position estimation algorithm and can be used for postprocessing position estimations and removing measurement noise. The proposed algorithm calculates the most probably path of a user, when taking into account historic movement traces. For some implementations it can be a useful add-on feature that improves the performance – especially in cases when modification and fine-tuning of a localization algorithm itself is not possible. The presented approach will work with any localization algorithm. Future work will focus on investigating the performance of the proposed algorithm using measured movement traces.

0.6

R EFERENCES

0.5

[1] F. Gustafsson and F. Gunnarsson, “Mobile positioning using wireless networks: possibilities and fundamental limitations based on available wireless network measurements,” Signal Processing Magazine, IEEE, vol. 22, no. 4, pp. 41–53, July 2005. [2] A. Sayed, A. Tarighat, and N. Khajehnouri, “Network-based wireless location: challenges faced in developing techniques for accurate wireless location information,” Signal Processing Magazine, IEEE, vol. 22, no. 4, pp. 24–40, July 2005. [3] A. Nicholson and B. Noble, “Breadcrumbs: Forecasting mobile connectivity,” in Proceedings of the 14th ACM international conference on Mobile computing and networking. ACM, 2008, pp. 46–57. [4] I. Akyildiz and W. Wang, “The predictive user mobility profile framework for wireless multimedia networks,” IEEE/ACM Transactions on Networking (TON), vol. 12, no. 6, pp. 1021–1035, 2004. [5] J. Petzold, F. Bagci, W. Trumler, and T. Ungerer, “Comparison of different methods for next location prediction,” Euro-Par 2006 Parallel Processing, pp. 909–918, 2006. [6] D. Ashbrook and T. Starner, “Using gps to learn significant locations and predict movement across multiple users,” Personal and Ubiquitous Computing, vol. 7, no. 5, pp. 275–286, 2003. [7] F. Lassabe, D. Charlet, P. Canalda, P. Chatonnay, and F. Spies, “Predictive mobility models based on kth markov models,” in IEEE int conf on pervasive services, 2006, pp. 303–306. [8] H. Lee, M. Wicke, B. Kusy, and L. Guibas, “Localization of mobile users using trajectory matching,” in Proceedings of the first ACM international workshop on Mobile entity localization and tracking in GPS-less environments. ACM, 2008, pp. 123–128. [9] F. Lassabe, P. Canalda, P. Chatonnay, and D. Charlet, “Refining wifi indoor positioning renders pertinent deploying location-based multimedia guide,” 2006. [10] P. Bahl and V. Padmanabhan, “Radar: An in-building rf-based user location and tracking system,” in INFOCOM 2000. Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, vol. 2. Ieee, 2000, pp. 775–784. [11] L. Rabiner and B. Juang, “An introduction to hidden markov models,” ASSP Magazine, IEEE, vol. 3, no. 1, pp. 4–16, 1986. [12] P. Hart, N. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,” Systems Science and Cybernetics, IEEE Transactions on, vol. 4, no. 2, pp. 100–107, 1968.

0.4 0.3 0.2 0.1 0

Straight line

90º turn

180º turn

Circular turn

(b) Results Fig. 8. Evaluation results for moving average and different Viterbi-based algorithms for the fundamental movements: straight line (A), 90 turn (B), 180 turn (C), circle (D).

The offline learning Viterbi algorithm has been trained with 300 sample trajectories using a forgetting factor wf = 0.005. Filter (1) Moving average (2) Viterbi - any direction (3) Viterbi - prefer linear (4) Viterbi - learning

Error offline 1.12 0.67 0.67 0.56

Error online 1.12 1.13 0.99 0.78

TABLE I R ESULTS FOR DIFFERENT ALGORITHMS . T HE UNIT IS GRID LENGTH .

Comparing first the offline and online modes for the Viterbi based algorithms it is clear that the experienced error is much lower when all measurements are available on beforehand. However for the moving average filter, the implementation is the same for the online and offline modes, resulting in similar performance. In offline mode, the Viterbi algorithms are significantly better than the moving average filter, however in online mode only the linear Viterbi filter and in particular the learning Viterbi filter outperform the moving average filter. This allows us to conclude that for online operation it is indeed worth using a learning Viterbi filter. For offline operation the two other Viterbi based filters perform nearly as well as the learning filter and still far better than moving average.

ICT–248894 WHERE2

A.16

D2.6

Directional Hidden Markov Model for Indoor Tracking of Mobile Users and Realistic Case Study

Jimmy Jessen Nielsen, Nicolas Amiot, Tatiana K. Madsen In Proceedings of The 19th European Wireless Conference (EW2013), Guildford, UK, April 2013. c

2013 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE.

140 / 146

Directional Hidden Markov Model for Indoor Tracking of Mobile Users and Realistic Case Study Jimmy Jessen Nielsen⇤ , Nicolas Amiot† , and Tatiana K. Madsen⇤ [email protected], [email protected], [email protected]

⇤ Networking

and Security, Department of Electronic Systems, Aalborg University, Denmark † IETR, University of Rennes 1, Rennes, France

Abstract—Indoors, mobile users tend to exhibit some level of determinism in their movement patterns during a day, for example when arriving to their office, going for coffee, going for lunch break, picking up print outs, etc. In this work we exploit this determinism to improve the accuracy of indoor localization systems. We consider two Hidden Markov Model (HMM) based filtering algorithms that use previous observations to estimate a user’s most likely movement trajectory, given a sequence of inaccurate location coordinates. The proposed Directional HMM algorithm is able to learn user habits by discriminating between different movement directions when populating the state transition probability matrix from training data. The proposed algorithm is compared to a Standard HMM algorithm that does not distinguish different movement directions. Evaluation results for a simple test scenario with two oppositely intersecting trajectories demonstrated a significant improvement of location accuracy with the Directional HMM algorithm. Further results for a scenario with realistic simulation based movement trajectories also showed improvements for 60% of the cases, however only if the HMM models are trained with usually unknown true trajectories. When trained with inaccurate location estimations, the HMM based algorithms showed no benefit compared to just using the localization system.

I. I NTRODUCTION Indoor localization of mobile users usually exploits Received Signal Strength (RSS) or Time-of-Arrival (ToA)/TimeDifference-of-Arrival (TDoA) measurements of available wireless networks to estimate distances between mobile users and anchor nodes. These distance estimates are in turn used together with geometrical techniques such as trilateration, triangulation, etc. to estimate the geographical coordinates of users. Such systems will inevitably experience some level of localization errors due to the impact of walls, furniture and other obstructions on the radio propagation. Even with additional filtering techniques such as Kalman and particle filters described in for example [1], [2], the localization process is rarely completely accurate. Indoors, users are often repeating the same movements, such as walking between main door, office, rest rooms, canteen, printer, and colleagues’ offices. State of the art localization approaches such as the Kalman and particle filters do not take this habitual determinism of the user into account. The main aim of the work described in this paper is to propose a method This work has been performed in the framework of the ICT project ICT248894 WHERE2, which is partly funded by the European Union.

for accounting for a user’s repeated behavior in the location determination process. Previously in [3] we proposed a similar algorithm, which however only supported movements between neighboring grid points and was only evaluated using simple and not very realistic movement traces. The main contributions of the present paper are 1) a direction-aware HMM based location tracking filter that allows to account for habitual movement patterns, and 2) a set of highly realistic movement traces for algorithm evaluation. In the literature there are several examples of existing work where learning movement patterns is exploited. Some examples are listed in the following: BreadCrumbs [4] tracks the movement of the device’s owner using GPS, and customizes a predictive mobility model for that specific user. The learned mobility model is a 2nd order Markov Chain, where each state represents a 2D geographical grid location covering 110m ⇥ 80m. Additionally, also the previous location is a part of the state information. The purpose of the protocol is to predict connectivity options a number of steps into the future, by combining Markov Chain probabilities and expected bandwith in those states. The authors in [5] compare different methods for prediction a user’s next location in an indoor office environment. Specifically, they compare dynamic Bayesian network, multilayer perceptron, Elman net, Markov predictor, and state predictor. For evaluation they use the Augsburg Indoor Location Tracking Benchmarks. This work did however not specifically consider detailed movement trajectories, but only larger scale locations such as rooms. Markov Models of different orders, where a number of previous states are taken into account, are used for movement prediction in [6]. The models are shown to work well in predicting the sequence of states that defines a user’s movements, with two sets of measurements, namely Augsburg Indoor Location Tracking Benchmarks and Nokia Context Data. An extended version of the fingerprinting based system RADAR proposed in [7] is proposed in [8]. Specifically, the authors have added information about the possible movements of a user such as wall restrictions. This is shown to slightly improve localization performance. The information about walls must however be provided a priori. Compared to the algorithm proposed in this paper, the work presented in [4], [5], [6] considers the prediction of user

movements on larger scales, i.e., resolutions measured in tens or hundreds of meters. The work presented in [8] exploits a priori environment knowledge, but does not use learned user behaviors as in the present paper. In this paper we propose a method for mobility learning and subsequent filtering of location information improving tracking of mobile users indoors that is based on the Hidden Markov Model (HMM) framework, see [9]. First, in section II we describe the considered radio-location based system as well as the HMM based modeling framework and the proposed extensions that enable mobility learning. For assessment of the proposed extensions, we present first a simple test scenario and results in section III, which is followed by a realistic case study in section IV. Finally, the conclusions and outlooks are given sin section V. II. S YSTEM M ODEL The considered system is a 2D geographical area Z, in which one or more users move around. The n’th user’s location at time t is a coordinate (xn (t), yn (t)), which however is not known. The users are instead being tracked by a radio-based localization system, which gives the location of the n-th user at time t as a coordinate (ˆ xn (t), yˆn (t)). The location of users is being estimated periodically with t time between samples. Further, we divide the geographical area into H ⇥ W grid points, meaning that a coordinate (ˆ xn (t), yˆn (t)) is represented by the grid point whose center is closest. Over time, a user’s movement history is represented by a discrete sequence of grid points corresponding to the location estimates obtained from the localization system. Depending on the movement speed and the location sample interval, the grid points in a trajectory can be neighbor points or points spaced farther apart. The used grid resolution will naturally set a limit on the achievable accuracy, which should be taken into consideration when evaluating the performance of the system. In practice, the users’ location estimates are already inaccurate due to for example radio propagation effects caused by nearby objects and obstructions. The inaccuracy due to the grid discretization may therefore not be significant compared to the error of the localization system. Given the discrete time and quantized grid point representation of user movements, we use the Hidden Markov Model (HMM) framework from [9] to facilite the considered trajectory filtering algorithm. In this work we consider two algorithms: 1) an HMM based algorithm where each state corresponds to a geographical location, denoted Standard HMM, and 2) an HMM based model with extended state space where each geographical grid point is represented by 9 states that encode the direction of travel, denoted Directional HMM. This encoding is elaborated after introducing the HMM framework. A. Hidden Markov Model Framework Definitions A HMM is characterized by the following [9]: • S = {S1 , S2 , S3 , . . . , SN } is the set of possible states, where N = W · H. Initially, for the Standard HMM

•

• •

•

•

algorithm we use a 1-to-1 mapping between states and geographical grid points in Z. A = {ai,j } is the state transition probability distribution between states i and j, where ai,j = P [qt+1 = Sj |qt = Si ], 1  i, j  N and qt is the actual state at time t. ⇡ = {⇡i } is the initial state probability distribution, where ⇡i = P [q1 = Si ]. V = {v1 , v2 , v3 , . . . , vM } is the set of possible observation symbols, which in our case corresponds to the set of grid locations. In this work the observable symbols correspond to the possible states, i.e., V = S. O = {O1 O2 O3 . . . OT } is a sequence of observations, where each Ot takes the value of a possible state in S. In our case O is a sequence of measured geographical grid locations made available from the localization system. B = {bj (k)} is the observation symbol probability distribution in state j, where bj (k) = P [vk at t|qt = Sj ], 1  j  N, 1  k  M.

B. Directional Hidden Markov Model

In order to support the learning of user habits, we consider also an extended state space where each geographical grid point (corresponding to a state in S) is represented by 9 states that encode from which direction the user entered the geographical grid point. This means that the number of possible states in S becomes 9 times larger. The direction is converted to a substate ID as shown in Fig. 1. The angle (in radians) is calculated using the atan2(y,x)1 function given the vector coming from subtracting two consecutive location coordinates. If a user has not moved to a different grid point between two consecutive samples, the direction will be encoded as substate 9 meaning idle. y π/2

8 π -π

7

6

1

5 2

3

4

x 0

-π/2

Fig. 1. Conversion of the movement direction into eight directional substates (innermost numbers that form a circle). The angle given from the atan2(y,x)) function is a value on the closed interval [ ⇡, ⇡]. In addition to the shown eight directions, substate 9 is used to indicate idleness.

C. Populating the Transition Probability Matrix A Depending on the considered algorithm Standard HMM or Directional HMM, the A matrix will have dimensions N ⇥ N or 9 · N ⇥ 9 · N , respectively. For both algorithms, the 1 This is a standard trigonometric function that computes the principal value of the arc tangent of y / x, using the signs of both arguments to determine the quadrant of the return value. The function is available in most programming languages, such as MATLAB, C, C++, Java, .NET, etc.

transition probabilities in the matrix are computed based on one or more learning movement trajectories. These are similar to the sequence of observations O that will be used later in the Viterbi algorithm. For every two consecutive coordinates in the trajectory, the state ids are determined by quantizing the coordinates into grid points and using the corresponding state ID from S. For the Direction HMM algorithm it is necessary to also compute the direction of movement between the points as specified in Fig. 1 in addition to just the grid points, in order to determine the appropriate state ids. The resulting list train train of training observations is Otrain = {otrain 1 , o2 , . . . , oT }. First, the matrix A is initialized with ones in all places, which means that transitions between all states will be possible, however unlikely. If this initialization is not used, the HMM has a tendency of getting stuck if there are no possible transitions towards the observed location. From the list of observations, the number of each type of transition is counted as follows: = 1 + ai=otrain , ai=otrain ,j=otrain ,j=otrain t t t 1 t 1

t = 2 . . . T.

(1)

E. Viterbi Algorithm for Trajectory Filtering Initially we consider simply the problem of determining the most likely path Q = {q1 q2 q3 . . . qT }, defined as a sequence of states from S, given a sequence of observations O = {O1 O2 O3 . . . OT }. This problem can be solved using the Viterbi algorithm [9], which has the steps described in the following. We have introduced a slight change compared to the formulation in [9], in the sense that we use separate indices i0 and j 0 to index B, which are equal the grid point indices of states i and j, respectively. For the Standard HMM algorithm i = i0 and j = j 0 , but for the Directional HMM algorithm a state ID conversion is necessary. 1) Initialization: 1 (i) 1 (i)

= ⇡i bi0 (O1 ), 1  i  N 0 = 0.

t (j)

= max 0 [ 1jN

t 1 (i) ai,j ] bj 0 (Ot ),

j = 1 . . . N 0,

(2)

t (j)

= arg max[ 1jN 0

2 t  T (7)

t 1 (i) ai,j ],

i=1

where N = N or N = 9 · N depending on the considered algorithm. 0

1 t  T (6) 1 i  N 0

0

ai,j = 1,

(5)

2) Recursion:

Finally, the matrix A is row-normalized so that: N X

(4)

0

1 i  N 0 . 3) Termination: qT⇤ = arg max[

D. Construction of Observation Probability Matrix B The B matrix specifies the probability of observing the different observation symbols when being in different states. In this work, this corresponds to the probability of getting a location estimate (ˆ x, yˆ) when being in location (x, y). Intuitively, this depends very much on the used localization system, as this determines how far away from the true location an location estimate is. A common way of modeling the location error of a localization system is to use a 2D Gaussian distribution. In this work we will as well make this assumption, as well as assuming that location estimates are unbiased meaning that the Gaussian distribution is zero-mean and described solely through a single standard deviation term xy . Hence, for each row in the B matrix, the symbol observation probabilities are calculated as a quantized 2D Gaussian distribution, depending on the distance between the state that the row corresponds to and each symbol. Further, the matrix B is row-normalized so that: N X bj (k) = 1, k = 1 . . . M. (3) j=1

Notice that the number of states in B is always N as we do not discriminate between directions for the observation probabilities. This could however be considered for a future extension, as it would allow to account for knowledge about walls or obstructions, by rating movements through a wall highly unlikely or impossible.

1iN 0

(8)

T (i)].

4) Path (state sequence) backtracking: qt⇤ =

⇤ t+1 (qt+1 ),

t=T

1, T

2, · · · , 1,

(9)

where ⇡i is the initial state probability for state i; t (i) is the best score (highest probability) along a single path, at time t, which accounts for the first t observations and ends in state Si ; t (j) is an array, which is used to keep track of the argument, i.e., the value of i, that maximized eq. (6); and qt⇤ is the state at time t for the best state sequence. III. R ESULTS : S IMPLE T EST S CENARIO Initially, we present results of the algorithms’ performance for a simple test scenario. This scenario is outlined in Fig. 2 and it consists of two base line trajectories (shown in black) and corresponding noise realizations. These represent the estimated trajectories that a location system would provide. Notice that the error term produced here is arbitrary and its effect should not be interpreted quantitatively. For the following results, xy = 2 m was used to generate the observation probability matrix B. This particular value is not necessarily an optimal choice as it was determined by manually inspecting the algorithms’ performance with a few different choices. We show results for using both the true baseline trajectories and observed (erroneous) trajectories to train the HMM, i.e., to populate the state transition probability matrix A. For the latter, 5000 noise realizations were used for

18

are based on 50 runs (25 for each baseline trajectory) of the HMM algorithms, where each run uses a noise realization that is not used for training.

16 14 12 10 8

Empirical CDF

1

6

0.8

4 2

0.6 5

10

15

20

25

F(x)

0 0

0.4

Fig. 2. A plot of 100 noise realizations of each of the two base-line test trajectories (shown in black). Notice that the two trajectories both start from the bottom of the plot, meaning that they meet from opposite directions. This should give the Directional HMM algorithm an advantage since it encodes the movement direction.

Localization system Loc. Sys. + std. HMM Loc. Sys. + dir. HMM

0.2

0 0

0.5

1

1.5

2 2.5 RMSE [m]

3

3.5

4

(a) Training: baseline trajectory.

the training. Fig. 3 shows two examples of the algorithms’ performance. Here we see that training with the baseline trajectory gives a trajectory closer to the true trajectory than if training with observation data.

0.8

F(x)

0.6

Baseline (true) trajectory Best guess trajectory Test trajectory

0.4

19 18 17 16

Localization system Loc. Sys. + std. HMM Loc. Sys. + dir. HMM

0.2

0 0

15 14 13 12

0.5

1

1.5

2 2.5 RMSE [m]

3

3.5

4

4.5

(b) Training: 5000 noise realizations.

11 10 9 8

Fig. 4.

7 6 5 4 3 2 1 0

Empirical CDF

1

0 1

2 3 4 5

6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

(a) Training: baseline trajectory. Baseline (true) trajectory Best guess trajectory Test trajectory 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

(b) Training: 5000 noise realizations. Fig. 3. Tracking performance of the Directional HMM algorithm for an example trajectory.

This observed behavior is further supported in the summarized results in Fig. 4. These plots of the Cumulative Distribution Function (CDF) of the Root-Mean-Square Error (RMSE)

Summarized tracking performance as CDF of RMSE over 50 runs.

Furthermore, the CDF plots reveal that the Directional HMM algorithm is more accurate than the Standard HMM algorithm. For example in Fig. 4(a) at the 90th percentile, the Directional HMM algorithm has an RMSE of only 0.5 m where the Standard HMM algorithm has an RMSE of 1 m. This difference is most likely because of the Directional algorithm’s ability to distinguish between directions. IV. R ESULTS : R EALISTIC C ASE S TUDY In this part we evaluate the proposed algorithms using realistic movement traces generated with the PyLayers simulator [10]. PyLayers is an open source radio simulator based on a graph description. It is designed to simulate complete dynamic scenarios with realistic movement of users inside a building, the transmission channel estimation for multiple radio access technologies, and the position estimation of users. The simulator can be seen as a stack of four layers: • The layout layer, which describes the simulation scene. • The mechanical layer, which simulates the human-like movement of the agents. • The network layer, which simulates the radio link connectivity between moving agents and anchors nodes. • The localization layer, which estimates the position of the moving agents. The first step of simulation is to initialize the layout layer with a building layout. From that layout, several graphs are built to describe the simulation scene in terms of outlines, used

coordinates [x, y] [2.32, 6.42] [14.37, 5.12] [18.3, 11.92] [40.02, 6.32]

materials and radio propagation paths. A complete description of those layout graphs is available in [11]. The mechanical layer is setup to introduce agents. Agents are the nodes that move through the Layout. Each agent independently chooses its destination and manages its movements in the layout, which includes avoiding walls and passing through the doors. Choosing a destination consists in selecting a room and finding an associated path to reach it. That path is a succession of rooms from the current position to the destination. It is obtained with the help of a graph from the layout layer. Once the path has been computed, the agents must reach their destination by taking into account the layout outlines. In order to obtain a human like mobility, the movement is modeled using magnetic forces. The magnetic force model supposes that the agents and the layout environment can be assimilated to positive poles and the destination to a negative pole. Hence, at each time instant, each agent is under the influence of several magnetic forces: an attractive magnetic force which attracts the agent to its destination and several repulsive magnetic forces from the walls to avoid the collision. From the network layer side, each moving agent is seen as a radio node which can communicate with other agents or static anchor nodes. All radio links are computed using a built-in ray-tracing tool. This gives the received power and time of arrival observations for each link. For the considered scenario, the agent position is estimated with the help of ToA observations. Those ToAs are environment-dependent simplified versions of what could be expected from an Ultra Wide Band (UWB) system. A maximum likelihood estimator described in [12] is used to obtained the position from those ToAs. 

  

 



Fig. 5. The five green dots represents the mobile agents. Each agent obtains its radio observations from the four anchor nodes, shown with red squares.

The considered test scenario is illustrated in Fig.5. The four anchors’ positions are given in Table I. The five moving agent have randomly chosen their destinations among all the reachable rooms of the layout. The simulation has been run for 8 hours, and the true positions and the estimated positions have been recorded every second for each agent.

Empirical CDF

1 0.8

0.6 F(x)

Anchor A1 A2 A3 A4

For the performance evaluations presented in the following, trajectories of nodes 2 5 have been used as training data to populate the transition probability matrix A, and the trajectory for node 1 was used for testing. For evaluation the test trajectory has been divided into segments of 50 seconds each, and the HMM algorithms have been run on the first 570 of these segments, which amounts to approximately 8 hours. The B matrix was generated with a standard deviation xy = 2 m.

0.4 Localization system Loc. Sys. + std. HMM Loc. Sys. + dir. HMM

0.2

0 0

1

2

3 RMSE [m]

4

5

6

(a) Training: True locations. Empirical CDF

1 0.8

0.6 F(x)

TABLE I A NCHORS ’ POSITIONS

0.4 Localization system Loc. Sys. + std. HMM Loc. Sys. + dir. HMM

0.2

0 0

1

2

3 RMSE [m]

4

5

6

(b) Training: Estimated locations. Fig. 6.

Summarized tracking performance as CDF of RMSE over 50 runs.

The CDF results in Fig. 6(a) show that at the 60th percentile, the HMM based algorithms improve the location RMSE from approximately 1.2 m to 0.7 m. However, when considering the 90th percentile, the accuracy of the HMM based algorithms is slightly worse than just using the localization system. Also, there is no noticable difference between the two HMM based algorithms; the Directional HMM algorithm is even slightly worse than the Standard HMM at the 90th percentile. This suggests that there is not much habitual behavior in the used training and test data, and secondly we speculate that an even longer movement trace for training data could be useful. The latter is based on the thought that the Directional algorithm has 9 times more transition probability entries in the A matrix and might therefore need more training data. Looking at the corresponding results in Fig. 6(b), where imperfect location estimates have been used for training, the use of the HMM based algorithms does not give any benefit. This contradicts the results obtained in the simple test scenario in section III. We believe that this difference comes down to the simple test scenario data being statistically unbiased, whereas the realistic ray-tracing based simulations may very well incorporate location-dependent bias of the location estimation due to specific radio propagation characteristics.

Baseline (true) trajectory Best guess trajectory Test trajectory 11 10 9 8 7 6 5 4 3 2 1 0

0 1 2 3 4 5 6 7 8 9 10 1112 13 1415 1617 1819 20 2122 2324 2526 27 2829 3031 3233 34 35

(a) Standard HMM Baseline (true) trajectory Best guess trajectory Test trajectory 11 10 9 8 7 6 5 4 3 2 1 0

0 1 2 3 4 5 6 7 8 9 1011 1213 14 1516 1718 19 2021 2223 24 2526 2728 29 3031 3233 34 35

(b) Directional HMM Fig. 7. Example trajectory and tracking performance of the two HMM algorithms. Estimated locations were used as training data.

From the summarized CDF results in Fig. 6 for the realistic case study we see that the HMM based algorithms do not improve performance in all cases. However, if the HMM models are trained using true location traces, significant improvements could identified at the 60th percentile. From this we conclude that there is a quite large amount of situations in which we can benefit from using the HMM based algorithms. On the other hand, in the remaining situations the HMM based algorithms make worse estimations than the localization system, which negatively impacts the location accuracy. This is exemplified in Fig. 7 where the standard HMM algorithm follows the test trajectory more closely. One can also notice that the Directional HMM algorithm uses fewer points, suggesting that it has too few learned transition options in the A matrix, and could benefit from more learning data. V. C ONCLUSIONS AND O UTLOOK In this work we have proposed a Directional Hidden Markov Model (HMM) trajectory filtering algorithm for improving indoor location tracking by taking into account habits in users’ movement patterns. This was achieved by considering an extended state space for the HMM, where the previous movement direction is encoded as one of nine substates for each geographical grid point in the considered area. Evaluation results for a simple test scenario with two oppositely intersecting trajectories demonstrated a significant improvement of location accuracy with the Directional HMM algorithm. Further results for a scenario with realistic simulation based movement trajectories also showed improvements for 60% of the cases, however only if the HMM models are trained with usually unknown true trajectories.

When trained with inaccurate location estimations, the HMM based algorithms showed no benefit compared to just using the localization system. This much lower improvement for the realistic movement traces we believe is partly due to insufficient habitual behavior in the realistic movement traces, as waypoints are randomly and independently chosen for users. The method used for training the HMM model, i.e., populating the transition probability matrix, has been done using a heuristic method. The HMM framework description in [9] outlines methods for solving the HMM problem no. 3: determining the best model parameters, namely the transition probability matrix, the observation probability matrix and the initial state probability vector. Using such an approach would allow for improving the location accuracy. So far, the proposed algorithms have only been considered for offline post-processing. For online usage, it would be necessary to continually run the Viterbi algorithm. As the backtracking step of the Viterbi algorithm is quite computationally intensive, it would be necessary to investigate how this can be done in an efficient manner. R EFERENCES [1] F. Gustafsson and F. Gunnarsson, “Mobile positioning using wireless networks: possibilities and fundamental limitations based on available wireless network measurements,” Signal Processing Magazine, IEEE, vol. 22, no. 4, pp. 41–53, July 2005. [2] A. Sayed, A. Tarighat, and N. Khajehnouri, “Network-based wireless location: challenges faced in developing techniques for accurate wireless location information,” Signal Processing Magazine, IEEE, vol. 22, no. 4, pp. 24–40, July 2005. [3] T. Laursen, N. Pedersen, J. Nielsen, and T. Madsen, “Hidden markov model based mobility learning for improving indoor tracking of mobile users,” in Positioning Navigation and Communication (WPNC), 2012 9th Workshop on. IEEE, 2012, pp. 100–104. [4] A. Nicholson and B. Noble, “Breadcrumbs: Forecasting mobile connectivity,” in Proceedings of the 14th ACM international conference on Mobile computing and networking. ACM, 2008, pp. 46–57. [5] J. Petzold, F. Bagci, W. Trumler, and T. Ungerer, “Comparison of different methods for next location prediction,” Euro-Par 2006 Parallel Processing, pp. 909–918, 2006. [6] F. Lassabe, D. Charlet, P. Canalda, P. Chatonnay, and F. Spies, “Predictive mobility models based on kth markov models,” in IEEE int conf on pervasive services, 2006, pp. 303–306. [7] P. Bahl and V. Padmanabhan, “Radar: An in-building rf-based user location and tracking system,” in INFOCOM 2000. Nineteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, vol. 2. Ieee, 2000, pp. 775–784. [8] F. Lassabe, P. Canalda, P. Chatonnay, and D. Charlet, “Refining wifi indoor positioning renders pertinent deploying location-based multimedia guide,” in 20th International Conference on Advanced Information Networking and Applications - Volume 2 (AINA’06). IEEE Computer Society, 2006. [9] L. Rabiner and B. Juang, “An introduction to hidden markov models,” ASSP Magazine, IEEE, vol. 3, no. 1, pp. 4–16, 1986. [10] http://www.pylayers.org. [11] B. Uguen, N. Amiot, and M. Laaraiedh, “Exploiting the graph description of indoor layout for ray persistency modeling in moving channel,” in Antennas and Propagation (EUCAP), 2012 6th European Conference on, march 2012, pp. 30–34. [12] M. Laaraiedh, L. Yu, S. Avrillon, and B. Uguen, “Comparison of hybrid localization schemes using rssi, toa, and tdoa,” Wireless Conference 2011 - Sustainable Wireless Technologies (European Wireless), 11th European, pp. 1–5, april 2011.