International Workshop on Synergies in Communications and Localization. Program. Expert Panel Start time: Thu, 18 Jun, 2:45 pm

Transcription

1

2

3 SyCoLo International Workshop on Synergies in Communications and Localization Date: Thursday, 18 June 2009 Room: Seminar 3&4 Organizers: Ronald Raulefs, German Aerospace Center (DLR) Marco Luise, University of Pisa Simon Plass, Xcitec GmbH Program Keynote 1 Start time: Thu, 18 Jun, 8:15 am Moe Win (MIT) Positioning with GNSS and Comm. Systems Start time: Thu, 18 Jun, 9:00 am GNSS Positioning in Critical Scenarios: Hybrid Data Fusion with Communications Signals C. Mensing, S. Sand, A. Dammann (German Aerospace Center (DLR)) D. Park (CEO, Geospatial Research Centre (NZ)) Expert Panel Start time: Thu, 18 Jun, 2:45 pm Expert Panel, moderated by R. Raulefs, M. Luise, and S. Plass Fingerprinting and Positioning Algorithms Start time: Thu, 18 Jun, 4:00 pm High-Performance Indoor Localization with Full-Band GSM Fingerprints B. Denby (Université Pierre et Marie Curie), Y. Oussar, I. Ahriz, G. Dreyfus (Laboratoire dšélectronique, ESPCI, ParisTech) A Scheme for Indoor Localization through RF Profiling I. T. Haque, I. Nikolaidis, P. Gburzynski (University of Alberta) Joint Estimation of Position and Channel Propagation Model Parameters in a Bluetooth Network J. Rodas, C. J. Escudero (Universidade da Coruña) Solving the Source Localization Problem via Global Distance Continuation G. Destino, G. T. Freitas de Abreu (University of Oulu) On the Performance of Hybrid GPS/GSM Mobile Terminal Tracking C. Fritsche, A. Klein (TU Darmstadt) Assessing GPS Robustness in Presence of Communication Signals B. Motella (Istituto Superiore Mario Boella), S. Savasta, D. Margaria, F. Dovis (Politecnico di Torino) Improving the Performance of TOA Over Wireless Systems to Track Mobile Targets M. Ciurana, F. Barcelo-Arroyo, M. Llombart (UPC) Cooperative Positioning and Cognitive Radio Start time: Thu, 18 Jun, 10:50 am Cluster-Based Ranging for Accurate Localization in Wireless Sensor Neworks S. Sergi, F. Pancaldi, G. M. Vitetta (University of Modena e Reggio Emilia) Fundamental Performance Limits of TOA-Based Cooperative Localization M. Nicoli, D. Fontanella (Politecnico di Milano) Coexistence Strategies and Capacity Theorems of Interference Awareness Cognitive Radio N. Yi, Y. Ma, R. Tafazolli (University of Surrey) Performance Limits on Ranging with Cognitive Radio D. Dardari (University of Bologna), Y. Karisan, S. Gezici (Bilkent University), A. A. D Amico, U. Mengali (University of Pisa) Keynote: Creating Smart Positioning Solutions Start time: Thu, 18 Jun, 2:00 pm 119

4 Organizing Committee General Chair Hamid Akhavan, T-Mobile International AG General Vice Chair Paul Kühn, University of Stuttgart Technical Program Chair Gerhard Fettweis, TU Dresden Wojciech Kabacinski, Poznan University of Technology Heinrich Stüttgen, NEC Europe Ltd Tutorial Co-Chairs Andrzej Jajszczyk, AGH University of Science and Technology Krakow Moe Win, Massachusetts Institute of Technology Application Panels Co-Chairs Jörg Eberspächer, Technische Universität München Steve Weinstein, Communication Theory and Technology Consulting Workshop Co-Chairs Klaus David, University of Kassel Heather Yu, Huawei USA Student Travel Grant Chair Eduard Jorswieck, TU Dresden Marketing Chair Klaus-D. Kohrt Finance Chair Volker Schanz, VDE Information Technology Society Bruce Worthman, IEEE Communications Society Local Organization Peter Rost, TU Dresden Peter Neu, VDE ComSoc Project Management June Leach-Barnaby, IEEE Communications Society Gayle Weisman, IEEE Communications Society IEEE Communications Society Doug Zuckerman President, John M. Howell Executive Director Mark Karol Vice President Conferences, Byeong Gi Lee Vice President Member Relations, Sergio Benedetto Vice President Publications, Andrzej Jajszczyk Vice President Technical Activities, Bruce Worthman Manager Conferences, Finance & Administration June Leach-Barnaby Senior Manager, Meetings & Conferences Gayle Weisman Senior Manager, Meetings & Conferences Heather Ann Sweeney Associate Marketing Manager 4

5 Joint Estimation of Position and Channel Propagation Model Parameters in a Bluetooth Network Javier Rodas, Carlos J. Escudero Departamento de Electronica y Sistemas. Universidade da Coruña. La Coruña. Spain jrodas@udc.es, escudero@udc.es Abstract Wireless Sensor Networks are a promising solution for indoor location systems. However, many of these systems rely on algorithms that use parametric models of the channel propagation where the parameters can be time variant. This paper introduces a new technique based on a Bayesian filtering method that estimates network nodes positions at the same time that propagation model parameters are extracted. Experimental results show the location estimation improvement of the introduced technique. INTRODUCTION Wireless Sensor Networks (WSN) are widely recognized as the solution for indoor location systems [1]. These networks are complementary to the Global Navigation Satellite Systems (GNSS) [2], which are only valid for outdoor environments where there is Line Of Sight (LOS) with the satellites. WSN are composed by several network nodes with wireless communication capability to communicate between them. By using this capability, we can extract information about transmission parameters such as the Received Signal Strength Indicator (RSSI). This information can be used to estimate distances between transmitter and receiver nodes and, therefore, WSN are very suitable for location systems. A target node location is estimated by considering anchor nodes whose positions are known. To estimate the position of the target node it is necessary to obtain the distances to, at least, three anchor nodes. There exist many technologies of WSN as for example: WIFI, Bluetooth, Zigbee, UWB, etc. [1], that allow to cover wide areas with multiple low-cost and low-power nodes. In this paper, we have used a Bluetooth sensor network that considers the Received Signal Strength Indicator (RSSI) as a physical parameter to estimate distances between nodes. Generally, the level of any RSSI measurement varies randomly according to the environment circumstances (presence of obstacles, multipath, people moving around,... ). Since they cause important fluctuations, the location system accuracy will be reduced. Nevertheless, these variations can be modeled with small and large scale models [3] that represent statistically the origin and type of these variations. The RSSI ranging technique is based on the principle of signal attenuation with the distance. In the paper, we have used a classical propagation model, that tries to describe the variations of the signal versus the distance as follows ( ) d P L (d)[db] = P L (d 0 ) + 10nlog + X σl (1) where P L (d) and P L (d 0 ) are the power attenuations of the received signal at the distances d and d 0, respectively, d 0 is a distance of reference, n is the exponential path-loss and X σl represents the random noise with a normal distribution, with zero mean and σ L standard deviation. Note that the random power variation is represented by means of a lognormal model. Although there are several techniques that try to reduce the measurement variations using sensor networks [2], if the assumption of the path-loss, n, is wrong these systems do not work correctly. The exponential path-loss n value typically varies between 1 and 3 in indoor environments when there is a clear Line Of Sight (LOS), but it can suddenly change if the line of sight is blocked, this is, in the case of Non Line Of Sight (NLOS). This parameter can be empirically estimated from off-line field measurements at several distances in the environment. However, since its value can change in the future, a technique to compute the exponential path-loss is needed. This paper introduces a new method for position estimation with dynamic adaptation of the path-loss parameter when n is discrete. The new method is based in a Bayesian filtering technique [4], [5] that uses the RSSI measurements from an experimental Bluetooth network. This paper is organized as follows: Section I presents the problem, emphasizing the possible changes that could take place in the path-loss. Section II introduces the algorithm based on a particle filtering that will make a joint estimation of position and discrete channel parameters. Section III shows the experimental results that prove the advantages of the proposed models and algorithms. Finally, section IV is devoted to the conclusions. d 0 I. PROBLEM STATEMENT The power attenuation model defined in (1) is described by the path-loss n, that should be determined a priori from experimental measurements. The values of n are dependent of the environment conditions and if a wrong value is assumed, it

6 will affect drastically to the distance estimation and, therefore, to the accuracy of the position estimation. We have made in a 6 10 meters indoor scenario the following experiment: We took real RSSI measurements at several distances (from 1 to 9 meters) between two Bluetooth nodes (target and anchor nodes), during about four minutes on each of them. To tell the difference between the Line Of Sight (LOS) and Non Line Of Sight (NLOS) cases, we repeat the measuring putting an obstacle at 40 cm in front of the target node, to block the direct LOS. In particular we used Bluetooth v2.0 nodes: One node was connected to a laptop via USB, model AirCable Host XR with a 2 dbi antenna. And the other one was a Nokia mobile phone. In order to obtain the measurements we used the standard discovery capability of Bluetooth called Inquiry. The anchor node acted as a master, transmitting Inquiry packages, while the mobile phone (target node) acted as a slave, replying to these discovery packages. The master was configured to process Inquiry responses in Inquiry with RSSI mode [7]. When the master receives a Inquiry response from the mobile phone it obtain the Received Signal Strength Indicator (RSSI), that is, the signal power in dbm of the received response, that is used to estimate distances between nodes. Figure 1 shows the power attenuation that we have observed in the LOS case at different distances between master and slave nodes. Note that the solid green line represents the noiseless model defined in (1) when considering a path-loss, n, deduced from a linear regression. On the other hand, the figure 2 shows the NLOS case. As shown, the deduced values of the path-loss where n LOS 1.8 and n NLOS 0.02, for LOS and NLOS cases, respectively. In order to achieve a reliable location estimation algorithm, it is necessary to adapt the model to track changes in the channel conditions. There are several works that estimate channel parameters. In [8] the unknown propagation model parameters are deduced from mathematic formulation. The work in [9] proposes parametric propagation models as a feasible way to track the channel, and [10] uses a Bayesian filtering (particle filter) in order to estimate the channel model parameters. However this method are only design to estimate this parameters without taking into account the joint estimation with the target node position. The problem of the joint estimation of the path-loss and position is very complex because each estimator, one for position and other for path-loss, needs information from the other and it is possible to reach multiple solutions. Therefore, in order to take into account channel conditions and guarantee the algorithm convergence, we must reduce the freedom degrees in the path-loss parameter. Our work uses a similar idea to [11] and [12] where there considers transitions between different situations (LOS and NLOS) by using a two node Markov model, that takes into account the probability of LOS between transmitter and receiver. Out algorithm considers the estimation of the target node position and, at the same time, the conditions of LOS (n LOS 1.8) or NLOS (n NLOS 0) in a discrete manner. In order to Loss (db) Fig. 1. Fig. 2. Loss (db) Measurements Model n=1.801! L = Distance (m) Power loss versus distance for the Line Of Sight (LOS) case. Measurements Model n=0.0274! L = Distance (m) Power loss versus distance for the Non Line Of Sight (NLOS) case. achieve this joint estimation we introduce a new algorithm based on a particle filter. II. PARTICLE FILTER ALGORITHM A particle filter is a Monte Carlo (MC) method for implementing a recursive Bayesian filter [4]. It is based on a set of random samples, named particles, associated to different weights that represent a probability density function (pdf). Basically, the objective is to construct the a posteriori pdf recursively, p(s i (t) z(t)), where s i (t) is the state of the i-th particle and z(t) is the observation (i.e. anchor nodes observed RSSIs), that we have at the t instant. Since our objective is to make the joint estimation of the target node position and the path-loss between target and anchor nodes, it is necessary to consider this information in the state of each particle. That is, the state corresponding to the i-th particle will be composed by the following components: s i = {x i, y i, n i1,..., n ik }

7 where (x i, y i ) are the coordinates 1, and n ij are the path-loss estimated parameters corresponding to the channel between the i-th particle and the j-th anchor node, with K anchor nodes. We will imposed a constraint to the n ij parameters in order to consider only the cases of LOS and NLOS with values equal to n LOS = 1.8 and n NLOS = 0, respectively. Moreover, each particle has an associated weight w i (t) directly related to p(s i (t 1) z(t 1)) [16]. We will consider N p particles with a random initialization of their states and equal weights w i = 1/N p. The algorithm performs several consecutive iterations, and each of them is divided in the following steps: prediction, update, resampling and estimation. A. Prediction It computes a new state for each particle based on the dynamic model that explains how their position coordinates (x i, y i ) and n ij channel parameters must be updated. In order to simplify the analysis we use a simplified model that considers static positions where coordinated are updated in a random manner. A more complete dynamic model could be used as in [15]. As it is shown in the scheme of figure 3, n ij parameters can only vary between two values (n LOS and n NLOS ). Note that a switch between these two states is only allowed with a low probability p change. The updates are explained as follows: x i (t) = x i (t 1) + n x t y i (t) = y i (t 1) + n y t n ij (t 1) if ((n n > p change ) ) nn p n n ij (t) = LOS if change && ( n ij (t 1) == n NLOS ) nn p n NLOS if change && n ij (t 1) == n LOS n x N (0, σ pos ) n y N (0, σ pos ) n n U(0, 1) (2) where N (µ, σ) is a Gaussian distribution with µ mean and σ standard deviation, U(0, 1) is a Uniform distribution between 0 and 1, and t is the time interval between iterations. p change B. Update In that second stage, the particle weights are updated and normalized as follows [16]: w i (t) = w i (t 1)p(z(t) x i (t)) w i (t) w i (t) = Np j=1 w j(t) where w i (t) stands for the normalized weights. In the update process, the conditioned probability of the observations versus the state depends on the propagation model (1). Taking into account the noise Gaussianity, we obtain the following expression for the j-th anchor node: 1 p(z j (t) s i (t)) = exp ( (z j(t) P L,ij (d)) 2 ) σ L 2π 2σ 2 L where P L,ij (d) is defined by the propagation model (1), but applying the appropriate n ij parameters, according to the j-th anchor node: ( ) d P L,ij (d)[db] = P L (d 0 ) + 10n ij log Considering the K anchor nodes, we obtain: C. Resampling p(z(t) s i (t)) = K p(z j (t) s i (t)) j=1 To avoid degeneration problems in the particle system, when many of them have low weights after some iterations, new particles must be generated. To this end, N p new particles are generated by a random selection from the current ones, by using a probability based on their weights. Thus, the strongest particles (particles with highest weights) will tend to be replicated, while the weakest ones will tend to be erased. D. Estimation Finally, the parameter estimation of the j-th anchor node is computed by means of a weighted sum of the state information from all the particles. It is computed as follows: d 0 1 p change LOS p change NLOS 1 p change Fig. 3. Particle n ij switch schema between LOS and NLOS cases, based on its probability of change p change x(t) = y(t) = n j (t) = N p w i (t)x i (t) i=1 N p w i (t)y i (t) i=1 N p w i (t)n ij (t), i=1 j = 1,..., K 1 Without loss of generality, we are assuming that everything moves in the same plane, avoiding azimuthal component.

8 III. EXPERIMENTAL RESULTS In this section we are going to check the performance of the introduced particle filter (2). As we have said before, the algorithm should estimate simultaneously the position and the n j propagation model parameters for each anchor node. In the following experiments, the possible values of n j are limited to detect LOS (n LOS = 1.8) or NLOS (n NLOS = 0). These two values are realistic and typical of an indoor real scenario, as shown in figures 1 and 2. Other parameters σ pos = 0.3 meters, σ n = 0.5, σ L = 5 dbm and t = 1 second were used in all the experiments. We considered 4 anchor nodes at each corner of a room of 6 10 meters, taking RSSI measurements. Then, we try to locate a target node that is situated in the coordinate (4,5) meters. The σ L value is also realistic and typical for real indoor environments as shown in figures 1 and 2. P(e< error ) LOS 3 LOS, 1 NLOS 2 LOS, 2 NLOS 1 LOS, 3 NLOS error [m] Fig. 6. Cumulative Distribution Function (CDF) of the position prediction error when we lose the LOS with 1, 2 or 3 anchors nodes. P(e< error ) =2.0 (error=+0.2) =1.9 (error=+0.1) =1.8 (no error) =1.7 (error= 0.1) =1.6 (error= 0.2) =0.2 (error= 1.6) =0.1 (error= 1.7) =0.0 (error= 1.8) error [m] Fig. 4. Cumulative Distribution Error (CDF) of the position prediction error for 4 LOS estimations, when one n j anchor node parameter is wrong. n n estimation (p change =0.02) n estimation (p change =0.05) n estimation (p change =0.10) value decisor threshold Time (s) Fig. 5. Path-loss n j discrete prediction to detect the LOS or NLOS case in the changing anchor node. A first experiment tries to demonstrate the effect of considering a fixed value of n in our algorithm, that is, when dynamic changes between LOS and NLOS path-loss are not considered. Assuming a fixed value of n ij = n LOS = 1.8, for each i-th particle with regard to each j-th anchor node, the figure 4 shows the Cumulative Distribution Error (CDF) of the location algorithm, when the real path-loss,, is different to the fixed value considered in our algorithm. If the difference from the real and the considered path-loss is low (differences of ±0.2 around 1.8), the performance of the algorithm is not drastically reduced. However, as soon as this real value changes to a NLOS situation, due to a obstruction in the LOS path with regard to one anchor node, the algorithm fails to do the position estimation. This experiment shows the necessity of considering an adaptive estimation of the path-loss with regard to the anchor nodes. In a second experiment, the efficiency of the introduced particle filter model is tested when the LOS with regard to one of the anchor nodes is lost in the middle of the simulation, while the three remaining anchors stay in a LOS condition. The figure 5 shows the tracking speed of this model for the anchor that changes its condition from LOS to NLOS, when three p change values were used. As shown in this figure, a low p change probability is necessary in order to achieve a good n j estimation. As soon as we increase p change we obtain a faster convergence of the n j estimation, but more noisy. However, if the noise in the n j estimation is not too high, we can apply a decisor to decide if we are in a LOS or NLOS condition. We use a decisor threshold n threshold = n LOS n NLOS 2 In particular for these experiments n threshold = 0.9. We choose a p change = 0.10 because the n j estimation is always below or above the n threshold value, with a fast convergence. Finally, figure 6 shows the performance of our algorithm when the following four conditions are considered: 0, 1, 2 or 3 anchors are in NLOS case while the remaining stay in LOS case. Note that for the first and third experiments we took RSSI measurements for a long period of seconds in order to obtain smooth CDF curves.

9 The results show a position accuracy lower than 1 meter for the 85% of the time for the first two cases. Note that only anchor nodes with LOS contribute to obtain a good estimation, and the nodes in NLOS condition only contribute with a small amount of noise in the position estimation. On the other hand, in order to estimate a position we need at least three anchor nodes. However, our system achieves good results when we only have 2 anchor with LOS, since it discriminates values outside the room, avoiding the ambiguity produced in those cases. Moreover, when we only have one LOS anchor, the algorithm does not work very well but still achieves noisy estimations, since the freedom degrees of the positions are reduced. IV. CONCLUSIONS This paper has introduced a location system that makes a join estimation of positions and discrete model propagation parameters. In order to achieve this joint estimation we consider a particle filter that includes in the particle states the channel propagation model path-loss value. In order to have good convergence, it is necessary to consider a discrete number of channel conditions. In our case, we consider the LOS and NLOS situations. The main advantage of the introduced technique is the dynamic adaptation of the algorithm to the time variant channel propagation conditions, while a target node location is also estimated. Experimental results have shown the good performance of this algorithm even when we lose the line of sight with regard to several anchor nodes. [11] M. Klepal, R. Mathur, A. McGibney, D. Pesch, Incluence of People Shadowing on Optimal Deployment of WLAN Access Points, VTC04, Los Angeles, USA, vol. 6, pp , Sept [12] E. Lutz, D. Cygan, M. Dippold, F. Dolainsky, W. Papake, The Land Mobile Satelite Communication Channel - Recording, Statistics, and Channel Model, IEEE Transactions on Vehicular Technology, vol. 40, No. 2, 1991, pp , May [13] A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, A. A. F. Loureiro, Localization Systems for Wireless Sensor Networks, IEEE Wireless Communications, pp. 6-12, Dec [14] Xinrong Li, RSS-Based Location Estimation with Unknown Pathloss Model, Wireless Communications, IEEE Transactions on, vol.5, no.12, pp , Dec [15] J. Rodas, C. J. Escudero, Bayesian Filter for a Bluetooth Positioning System, IEEE International Symposium on Wireless Communication Systems (ISWCS), Reykjavik, Iceland, Oct [16] S. Arulampalam, S. Maskell, N. J. Gordon, and T. Clapp, A Tutorial on Particle Filters for On-line Non-linear/Non-Gaussian Bayesian Tracking, IEEE Transactions of Signal Processing, Vol. 50(2), pages , Feb ACKNOWLEDGMENT This work has been supported by: 07TIC019105PR (Xunta de Galicia, Spain) y TSI (Ministerio de Industria, Turismo y Comercio, Spain). REFERENCES [1] J. Hightower, G. Borriello, Location Systems for Ubiquitous Computing, IEEE Computer Magazine, pp , Aug [2] M. D. Goran and R. E. Richton, Geolocation and assisted GPS, IEEE Communications Magazine, pp , Feb [3] T. S. Rappaport, Wireless Communications: principles and practice, 2nd edition, Prentice Hall, [4] A. Doucet, S. Godsill, C. Adrieu, On Sequential Monte Carlo Sampling Methods for Bayesian Filtering, Statistics and Computing, no. 10, pp , [5] F. Dieter, Jeffrey Hightower, Lin Liao, Dirk Schulz, Gaetano Borriello, Bayesian Filtering for Location Estimation, IEEE Pervasive Computing, vol. 02, no. 3, pp , Jul-Sept [6] D. Moore, J. Leonard, D. Rus, and S. Teller, Robust distributed network localization with noisy range measurements, in Proceedings of the Second ACM Conference on Embedded Networked Sensor Systems (SenSys 04), Baltimore, MD, Nov [7] Bluetooth Special Interest Group (SIG), [8] X. Li, RSS-based Location Estimation with Unknown Pathloss Model, IEEE Transactions on Wireless Communications, vol. 5, issue 12, pp , Dec [9] P. Tarrio, Ana M. Bernardos, J. R. Casar, An RSS Localization Method Based on Parametric Channel Models, SensorComm, Valencia, Spain, Oct [10] J. Rodas, C. J. Escudero, Particle Filtering for Channel Propagation Model Parameters Estimation, (submitted to) IEEE 69th Vehicular Technology Conference (VTC), Barcelona, Spain, Apr