Detection and Tracking Using Particle-Filter-Based ... - IEEE Xplore

1332

IEEE TRANSACTIONS ON MOBILE COMPUTING,

VOL. 9,

NO. 9,

SEPTEMBER 2010

Detection and Tracking Using Particle-Filter-Based Wireless Sensor Networks Nadeem Ahmed, Member, IEEE, Mark Rutten, Travis Bessell, Salil S. Kanhere, Member, IEEE, Neil Gordon, Member, IEEE, and Sanjay Jha, Senior Member, IEEE Abstract—The work reported in this paper investigates the performance of the Particle Filter (PF) algorithm for tracking a moving object using a wireless sensor network (WSN). It is well known that the PF is particularly well suited for use in target tracking applications. However, a comprehensive analysis on the effect of various design and calibration parameters on the accuracy of the PF has been overlooked. This paper outlines the results from such a study. In particular, we evaluate the effect of various design parameters (such as the number of deployed nodes, number of generated particles, and sampling interval) and calibration parameters (such as the gain, path loss factor, noise variations, and nonlinearity constant) on the tracking accuracy and computation time of the particle-filter-based tracking system. Based on our analysis, we present recommendations on suitable values for these parameters, which provide a reasonable trade-off between accuracy and complexity. We also analyze the theoretical Crame´r-Rao Bound as the benchmark for the best possible tracking performance and demonstrate that the results from our simulations closely match the theoretical bound. In this paper, we also propose a novel technique for calibrating off-the-shelf sensor devices. We implement the tracking system on a real sensor network and demonstrate its accuracy in detecting and tracking a moving object in a variety of scenarios. To the best of our knowledge, this is the first time that empirical results from a PF-based tracking system with off-the-shelf WSN devices have been reported. Finally, we also present simple albeit important building blocks that are essential for field deployment of such a system. Index Terms—Wireless sensor networks, simulations, experiments, performance attributes, measurements.

Ç 1

INTRODUCTION

W

IRELESS Sensor Networks (WSN) are increasingly being used in a variety of applications ranging from environmental monitoring to industrial automation. A particularly promising military application involves using WSN for detecting and tracking moving targets such as tanks, vehicles, and troops. Detection and tracking of targets is a mature and well-established research area. However, the current solutions rely on expensive and bulky sensors. The use of low-cost sensor nodes is an attractive and complementary approach. Among several tracking algorithms in the literature that use nonlinear filters, the Particle Filter (PF) [1] has been a popular choice. The PF (also known as the sequential Monte Carlo method) approximates a belief state for the presence of the target by means of many but finite random samples. Previous work [2], [3], [4] has demonstrated that the PF can be effectively employed in WSN systems for tracking applications. The tracking and detection performance of such a system is significantly influenced by several design

. N. Ahmed, S.S. Kanhere, and S. Jha are with the School of Computer Science and Engineering, University of New South Wales, Building K17, Anzac Parade, Kensington, 2052 Sydney, Australia. E-mail: {nahmed, salilk, sanjay}@cse.unsw.edu.au. . M. Rutten, T. Bessell, and N. Gordon are with the Defence Science and Technology Organization (DSTO), 200 Labs, PO Box 1500, Edinburgh, SA 5111, Australia. E-mail: {mark.rutten, travis.bessell, neil.gordon}@dsto.defence.gov.au. Manuscript received 19 Dec. 2008; revised 2 July 2009; accepted 15 Dec. 2009; published online 22 Apr. 2010. For information on obtaining reprints of this article, please send e-mail to: [email protected], and reference IEEECS Log Number TMC-2008-12-0503. Digital Object Identifier no. 10.1109/TMC.2010.83. 1536-1233/10/$26.00 ß 2010 IEEE

parameters (such as number of sensors, number of generated particles, and sampling frequency) and estimation of the calibration parameters (such as sensor gain, path loss factor, noise, and nonlinearity constant). However, all prior work simply assumes a particular set of values for these parameters without providing any insight into how they affect the behavior of the tracking system. In this paper, we conduct extensive simulations in a realistic environment to study the effect of the aforementioned parameters on the tracking and detection performance of the system. We consider the problem of simultaneous detection and tracking of an object moving through a particular target region. In our system, sensor nodes equipped with acoustic sensors are deployed in the target region. The sensors periodically sample the ambient sound and relay the measured samples to a central base station. The moving object generates an acoustic signature as it travels, which is captured by the acoustic measurements of the motes that are located near the object. The base station runs the PF algorithm on the collected samples, with the intensity of the object’s acoustic signal acting as the input. The PF estimates the presence of the object and, if the object is declared present, also estimates the object’s trajectory. In this work, we have conducted extensive simulations of this tracking system using NS2 to study the impact of various design and calibration parameters on the tracking performance. The two metrics that we have used to evaluate the performance of the system are: track estimation error, i.e., the distance between the estimated and the actual location of the target and computation time, i.e., the time required by the PF to execute. Our simulation environment takes into account Published by the IEEE CS, CASS, ComSoc, IES, & SPS

AHMED ET AL.: DETECTION AND TRACKING USING PARTICLE-FILTER-BASED WIRELESS SENSOR NETWORKS

several real-world effects such as wireless channel propagation and network protocol behavior such as delay and packet losses due to collisions. Based on our observations, we suggest suitable values for the relevant design and calibration parameters. To determine the benchmark for the optimal tracking performance, we analyze the theoretical Crame´r-Rao Bound [5]. The results of our simulated PF algorithm closely match this theoretical bound. We have also made significant experimental contributions by developing a prototype of our tracking system using off-the-shelf Xbow MicaZ motes. First, we conduct extensive experiments to calibrate the microphones of the motes. This includes characterizing the gain, path loss, noise variance, and the nonlinearity constant, which collectively form the calibration parameters. These calibration values are used in our MATLAB implementation of the PF algorithm. Second, we present empirical results from a real-world implementation of our tracking system. The results from our experiments are promising and demonstrate that our system can successfully track and detect an object moving along a variety of trajectories. Third, we summarize several system challenges encountered in our implementation and describe the solutions adopted for overcoming these challenges. Our system-level contributions include custom algorithms for high frequency acoustic sampling, time synchronization, shared channel access using Time Division Multiple Access (TDMA), and clustering. This paper makes the following specific contributions: We present a detailed simulation study to evaluate the effect of design and calibration parameters on the performance of a particle-filter-based WSN tracking system. The simulation results conform with the theoretical performance bound. . We propose a simple yet effective calibration mechanism for characterizing the behavior of offthe-shelf sensor hardware. This calibration mechanism is shown to perform better than other naive proposals (Section 4.1). . We implement a working prototype of our system on a sensor network. Our system can accurately detect and track an object moving along several different trajectories. . We present several system design solutions that lower the technical barriers for field deployment of WSN tracking systems. The rest of this paper is organized as follows: Section 2 summarizes related work. In Section 3, we provide a brief overview of the PF and discuss how we adapt it to our system. Interested readers are referred to a book by one of the coauthors [1] for further elaboration. The simulation-based evaluations along with relevant discussions on appropriate design and calibration parameters are presented in Section 4. This section also includes a detailed discussion on calibrating the microphones of sensor nodes. Section 5 presents the derivation of the Crame´r-Rao bound and a comparison of our simulation results with this theoretical bound. Results from our experiments are presented in Section 6. Section 7 details the systems challenges faced during the experimentation and describes the steps taken to overcome them. Finally, Section 8 concludes the paper. .

2

1333

RELATED WORK

In this section, we present a brief overview of previous work on using miniature sensor devices for target tracking. The work introduced in [2] is one of the earliest attempts at using tiny acoustic sensor devices for tracking purposes. The target is estimated via triangulation, i.e., comparing the difference in sound propagation delays from the sound source to different acoustic sensors. In [3], Gu et al. developed a lightweight multimodal detection algorithm for mote level microsensors. They discovered that simple fusion algorithms such as moving averages with thresholds are useful in object detection using WSN. Unfortunately, both of these studies assume that the sensor readings are free of ambient noise, which is a highly unrealistic assumption. In a typical outdoor environment (especially given the hostile nature of battlefields), it is expected that the sensor readings would be significantly influenced by ambient noise. In addition, given that the sensor nodes are small form factor devices and of low cost, it is expected that the readings would inherently be noisy due to their low fidelity. In [4], Duarte and Hu evaluated different machine learning algorithms in the context of vehicle detection. The authors proposed a two level detection architecture to increase the reliability of the PF. Different target detection algorithms, such as K-nearest neighbor, maximum likelihood classifier, and support vector machine classifier, are evaluated at a local node level. Then, the results of the local node level evaluation are passed to a group, which is formed dynamically. The fusion algorithms are performed at the group level nodes. However, resource-intensive tasks such as Fast Fourier Transform (FFT) are required to be performed at the local level nodes. Therefore, it is not suited for low-cost sensors. Simon and coworkers designed and implemented a sniper localization system based on acoustic signal processing and triangulation in [6]. Special hardware (Digital Signal Processing board) was designed in [6] for the resource-intensive acoustic signal processing tasks. In [7], He et al. designed and implemented a WSN with magnetic, acoustic, and motion sensors, which could classify a moving target such as a walking person or a vehicle. The motion sensor used in this work is a micropower impulse radar. Due to their high cost (typically US$5K), they may not be a suitable choice for many WSN systems. Coates and Ing also made use of the PF for target tracking in [8]. However, they propose to model each mote as a particle. Consequently, the corresponding real-world deployment would require thousands of motes to achieve tracking performance comparable to their simulation results. Further, the authors assumed that the motes have a fixed sensing range of 8 meters, a fixed detection probability of 0.7 within this sensing range and also assumed absence of any communication errors. All these assumptions make it difficult to apply the proposed algorithm in a real-world system. Recent examples of tracking work using off-the-shelf WSN devices include [9], [10], and [11]. All of these solutions require the presence of a cooperative target to generate specific sequence of signals, i.e., RF and audio [9] or target equipped with inertial sensors [11]. Our technique is based on sampling acoustic signals only thus the system can be trained to track the movement of any target (e.g., vehicles) that generates acoustic signals in real life.

1334


Distributed implementation of PF and resampling algorithms has also been proposed in [12]. However, the focus was on distributed computing, and the communication overhead of the proposed approach is unreasonable for WSN applications. Similarly, work presented in [13] is based on linear dynamics and observations. Unfortunately, most of the target tracking applications are highly nonlinear systems. In the literature, most of the performance analysis on collaborative signal processing is conducted by simulations and theoretical analysis, concentrating on exploring the design space and trade-offs under specific constraints and assumptions. The constraints and assumptions typically simplify the complicated real-world environments, which make it difficult to observe identical results, when the corresponding algorithms are implemented in real-world systems. What is lacking is a comprehensive study of the PF, based on realistic assumptions, which explains what parameters have the most significant effect on the PF behavior and can be used as a guideline for its implementation. In this work, we have carried out a detailed evaluation of the effect on various system parameters on the performance of the PF in a realistic simulation environment. In addition, experiences with a real-world implementation of the PF are also included. To the best of our knowledge, both aforementioned contributions are the first of their kind in the sensor network research community.

3

OVERVIEW OF THE PARTICLE FILTER

In this section, we provide a brief description of the particle filter algorithm used in our work. Among the several variants of the PF that are available, we use a recursive Bayesian tracking algorithm referred to as the Track-BeforeDetect Particle Filter (TBD-PF) in [1], [14]. Using this type of filter allows the information in the measurements from all sensors to be incorporated exactly into the estimation of the target state. The PF is a suboptimal nonlinear filter that performs estimation using sequential Monte Carlo methods to represent the probability density function of the target state. Note that another popular choice for tracking, the Kalman Filter (KF), assumes that system and the measurement processes are linear. However, in a variety of real scenarios, the assumptions of KF do not hold and approximate techniques must be employed. The Extended KF (EKF) approximates the models used for the measurement process to approximate the probability density by a Gaussian. There are several advantages in using a PF-based estimator over other nonlinear filtering approaches such as the EKF. These include: . . .

.

Target presence and absence are explicitly modeled by the probability function. The method can track targets moving randomly in the field of deployment. Non-Gaussian noise in sensor readings can be incorporated into the filter by estimating the distribution function of this noise. This incorporates the noise due to calibration errors in sensors in addition to the environmental noise. It permits us to detect targets with variable levels of intensity.

VOL. 9,

NO. 9,

SEPTEMBER 2010

The design of the filter has been limited to estimating the state of a single target for the purposes of this work. We now give a step by step overview of the operation of the PF and discuss how it is adapted to meet the requirements of our system.

3.1 Target Model We begin by assuming that N sensors are deployed in an n m surveillance area, with known positions ðxi ; yi Þ, i 2 f1 . . . Ng. We assume that the sensors are equipped with microphones. The sensors periodically sample the ambient noise and transmit the readings to a central base station. Suppose that a target is moving in this area according to a known dynamic model: Xtþ1 ¼ F Xt þ vt ;

ð1Þ

where . . .

vt is the process noise, normally assumed to be Gaussian noise with covariance matrix Q. t is time index. Xt is the target state vector defined as Xt ¼ ½xt ; x_t ; yt ; y_t ; It T ;

ð2Þ

where (xt ; yt ) and (x_t ; y_t ) denote the position and the velocity of the target and It denotes the target’s acoustic intensity. The existence of the target in the data is modeled as a binary Markov process. The target existence variable, Et , can take on two values, namely Et ¼ 0 indicating the absence of the target and Et ¼ 1 denoting its presence. The target can appear at any place and at any time step. Following its appearance, the target proceeds on a trajectory until it disappears, i.e., the intensity of the target signal strength falls below the sensors’ sensitivity level. We can model the transitional probabilities of the target birth, Pb , and its death, Pd , as follows: Pb ¼ PrfEt ¼ 1jEt1 ¼ 0g; Pd ¼ PrfEt ¼ 0jEt1 ¼ 1g:

ð3Þ ð4Þ

It is assumed that these probabilities are known a priori. However, if they are not known a very low value is assumed (e.g., 0.01). The motion matrix, F , in (1) for a sampling interval of T is given by 2 3 1 T 0 0 0 60 1 0 0 07 6 7 7 F ¼6 ð5Þ 6 0 0 1 T 0 7: 40 0 0 1 05 0 0 0 0 1 The covariance matrix Q of vt is given by 2 T 3 q1 3 6 T 2 q1 6 2 6

Q¼6 0 6 4 0 0

T 2 q1 2

T q1 0 0 0

0 0

0 0

T 3 q1 3 T 2 q1 2

T 2 q1 2

0

T q1 0

3 0 7 0 7 7 ; 0 7 7 5 0 T q2

ð6Þ


where q1 and q2 control the uncertainty of the target trajectory in position and intensity, respectively.

3.2 Sensor Model Each sensor i 2 f1 . . . Ng provides an acoustic measurement at discrete instants of time, t. This measurement is made by calculating the variance (the mean acoustic power) of 1,000 acoustic samples taken at a constant sampling rate. It is assumed that each of these samples is approximately Gaussian distributed, with zero mean and variance given by 2i þ i It d t;i :

ð7Þ

The first term of (7), 2i , is the measurement noise variance for sensor i, which represents both internal sensor noise and constant background noise. The second term represents the part of the measured signal due to the target. The target is assumed to be emitting a white acoustic signal of constant power, It , i is the gain factor for sensor i, and is the path loss, assumed identical for each sensor. The distance dt;i from the target to sensor i at time t is given by qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð8Þ dt;i ¼ ðxi xt Þ2 þ ðyi yt Þ2 : It can be easily shown that the measurement calculated from the variance of the 1,000 Gaussian samples is 2 distributed with 1,000 degrees of freedom. This distribution can be quite accurately approximated by a Gaussian with mean and variance hi ðXt Þ ¼

2i

i It d t;i ;

þ 2 2 2 Ri ðXt Þ ¼ c i þ i It d t;i ; Ns

the target dynamics (1) and the weights are calculated using the likelihood function arising from the measurement equation (11) assuming that the target is present ðpÞ

wt /

1.

2.

3.

where wi ðXt Þ is Gaussian distributed with zero mean and variance Ri ðXt Þ and wi is Gaussian distributed with zero mean and variance N2s c4i . The complete measurement recorded at time t is denoted as Zt ¼ fzt;i : i ¼ 1 . . . Ng, and the set of all measurements up to time t is denoted Z1:t ¼ fZk : k ¼ 1 . . . tg.

pðXt jZ1:t Þ

P X

ðpÞ ðpÞ wt Xt Xt ;

ð13Þ

Particle proposal (birth): A set of PB particles, ðB;pÞ B , is generated assuming that there is no fXt gPp¼1 target in the data. In this case, we randomly place samples around the coverage region. Particle proposal (target): A set of PT particles, ðT ;pÞ T fXt gPp¼1 is proposed from the particles at the ðpÞ previous time, Xt1 , sampling from (1) in combination with the Rao-Blackwellized velocities. Weight calculation: Once we have placed all the particles we need to compute their associated weights using (13) and the measurement from the current time, Zt ðS;pÞ

~t w

¼

N Y

ðS;pÞ ðS;pÞ ; Ri Xt ; N Zt ; hi Xt

ð14Þ

i¼1

and then normalize ðS;pÞ

wt

4.

ðS;pÞ

¼

~t w

ðSÞ

Wt

;

where

ðSÞ

Wt

¼

PS X

ðS;iÞ

w ~t

;

; e e þ Pd Pt1 þ ð1 Pb Þð1 Pt1 Þ ðT Þ ðBÞ e e : ¼ Wt ð1 Pd ÞPt1 þ Wt Pb 1 Pt1

ð12Þ

p¼1

5.

ð15Þ

i¼1

where S can refer to either B or T . Probability of existence: The probability of a target existing in the data can be calculated using the sum of the weights of the birth and target particles [16] Pte

where ðÞ is the Dirac delta function. The particles and their weights are updated recursively as new measurements become available. At each time step, the particle positions are proposed from their previous position by sampling from

N fZt ; hi ðXt Þ; Ri ðXt Þg;

where N fx; ; 2 g is the multivariate Gaussian function with mean and variance 2 evaluated at x. Since, the target motion model is linear in the target state and the measurement model does not depend on the velocities a technique known as Rao-Blackwellization [1] or marginalization [15] can be used to update the velocities of each particle. In this case, each particle uses a standard Kalman filter to update the velocities exactly, given the positions, rather than rely on Monte Carlo methods to explore the velocity space. The method is one of a family of variance reduction techniques, which aim to reduce the variance of the particle weights resulting in a more efficient filter. At each time step, the particle filter implemented here forms two sets of particles. One set of particles carries information about the target from the previous time step (called the target particles), while the other set searches for a new target in the data (called the birth particles). The probability of the target existing in the data, Pte ¼ PrfEt ¼ 1g, can be calculated as a function of the weights of these two sets of particles [16]. The resulting particle filter algorithm follows:

ð10Þ

3.3 Estimation Algorithm The basic function of the particle filter is to approximate the posterior density of the target state, given all measureðpÞ ments, by a set of P points, Xt , called particles, and ðpÞ corresponding weights, w [1]. That is,

N Y i¼1

ð9Þ

where c is a constant used to model nonlinearities in the system and Ns ¼ 1;000 is the number of samples in the measurement. Section 4.1 discusses the method of calibration used to determine 2i and i for each sensor. The sensor model can, thus, be summarized as hi ðXt Þ þ wi ðXt Þ; if Et ¼ 1; ð11Þ zt;i ¼ wi ; otherwise;

1335

ð16Þ ð17Þ

Combining: The birth and target particles are then combined into one set of PB þ PT particles with weights

1336

IEEE TRANSACTIONS ON MOBILE COMPUTING, ðB;pÞ

wt

ðT ;pÞ wt

6.

7.

ðBÞ ðB;pÞ e Wt wt ; ¼ Pb 1 Pt1 e ðT Þ ðT ;pÞ ¼ 1 Pd Pt1 Wt wt ;

4

ð19Þ

followed by another normalization. Resampling: The last step in the PF is the resampling process. Resampling eliminates particles with weights that are of low importance and multiplies those with higher values. The resampling also reduces the set of PB þ PT particles down to PT , ðpÞ T all with resulting in the set of particles fXt gPp¼1 uniform weights. We use the systematic resampling procedure described in [1]. If Pe is above a predetermined threshold then a target is declared present, where the particles resulting from Step No. 6 describe the target state. An estimate of the target position can be calculated from the set of particles by taking their mean PT X ðpÞ ^t 1 X : X PT p¼1 t

8.

ð18Þ

ð20Þ

For the next time step, we collect a new set of readings Ztþ1 and go back to Step 1.

SIMULATION RESULTS

As discussed earlier, one of the goals of this study is to determine the effect of various parameters on the performance of the tracking system. A secondary goal is to provide recommendations on the choice of these design parameters to engineer a real WSN-based tracking solution. The simulations described in this document have been conducted using the NS2 discrete event simulator using realistic values of transmission and sensing ranges resulting in a multihop communication network. Preliminary simulation results appeared in a previous publication [17]. For our simulation studies, we assume that N sensors are statically deployed in an area of size 120 m 120 m. We studied two popular and commonly used deployment topologies: 1) Grid: where the nodes are placed at an equal distance from each other resembling a perfect grid shape and 2) Uniform Random (UR): where the nodes are uniformly distributed over the entire field of deployment. However, to avoid large areas without sensor coverage, in the latter case, we divided the area into a number of smaller cells of equal size (the number of equal sized cells depends on the total number of nodes that are being deployed) and then randomly placed a node in each of the resulting cells. We assume that each node is equipped with an acoustic sensor and that it samples the ambient sound at predefined intervals. We refer to these intervals as sampling intervals/ time steps in the rest of this paper. As the NS2 simulator lacks acoustic sensor model, we have modeled the acoustic samples measured by the sensors as a simple distance-based function, wherein the intensity of these readings is an inverse function of the euclidean distance between the current position of the target and the location of the sensors. The acoustic measurements recorded by the sensors are, therefore, not actual readings but rather synthetically generated based on the calibration results.

VOL. 9,

NO. 9,

SEPTEMBER 2010

TABLE 1 Simulation Parameters

For simplicity, we limit this work to estimation of a single target moving along a straight diagonal line from a point at the bottom left corner of the topology to the right hand top corner. The velocity of the target was set to 0.35 m/sec and each simulation was run for 450 seconds. The base station is located at the center of the topology (at location 60,60). The complete simulation consists of two parts. The first part is performed in NS2 where all the measurements during the simulation run are recorded by each node and then these measurements are forwarded to the centrally located base station, over multihop, where they act as inputs to the PF. The PF assumes the values q1 ¼ 0:002 and q2 ¼ 106 , in reference to (6), and low values for the birth and death probabilities, Pb ¼ Pd ¼ 0:01, in this case. Once all the data are available, the base station runs the PF code (offline, in MATLAB). The NS2 simulation parameters are listed in Table 1. For MAC, we used 802.11 with RTS and CTS turned off. To quantitatively measure the performance of the tracking system, we define a metric, accuracy of estimation as the average euclidean distance between the actual and estimated location of the target. This metric is computed at each step of the execution of the PF. The second metric used in our evaluation is the computation time required for executing the PF algorithm. The total computation time is Tcomp ¼ Tns þ Tpf where Tns is time (real time) taken in Ns2 to collect all the measurements at the base station for a particular run of simulation and Tpf denotes the time taken by MATLAB to run the PF at the base station. For each run of the simulation, the measurements were taken for each node at the sampling interval T . For each time step, when the PF detects the presence of the target, the error is calculated in terms of the euclidean distance between the estimated location as indicated by the PF and the actual target location. For each run of the simulation, we calculated the quadratic mean (Root Mean Square, sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi PN 2 i¼1 erri ; RMS ¼ N where N is the total number of samples) of these errors across all the sampling intervals. We repeated each of our simulations 100 times and the results are shown as an average of all RMS error values. We also calculate the standard deviation in RMS values for 100 runs of the simulations. A similar process is repeated for the second metric, the computation time.

4.1 Calibration Calibration of the nodes is an important aspect of a realistic target tracking system [18]. Recall that our system uses the acoustic samples recorded by the sensor microphones as the


TABLE 2 Summary of the Parameters Estimated for Each Sensor

input signal to the PF. In this section, we present a simple yet effective mechanism for calibrating the sensor microphones. We assume that a white noise source of unit (arbitrary) intensity is available, which can be placed at varying distances from the sensor to be calibrated. We have chosen a white noise source over a tone, since our targets of interest such as vehicles typically have a broadband acoustic signature. Further, a white noise source is more robust to environmental effects such as multipath reflections in an indoor environment. Referring to the mean of the measurement function (9), repeated here, dropping the time index for brevity zi ¼ 2i þ i Id i ;

ð21Þ

the mean (21) and variance (10), of the measurement function involve four calibration parameters, namely the gain i , the noise variance 2i , the path loss factor , and the constant c used to model nonlinearities in the system. In order to accurately track a target, the gain, i , and the noise variance, 2i , must be determined for each sensor i, while it is assumed that the path loss factor, , and the nonlinearity constant, c, are the same for each sensor. We assume that the resulting parameters are independent of the sensor orientation, which may or may not be valid, depending on the sensor and the sensing environment. From (21), if there is no signal from the source (I ¼ 0 or di ! 1) then each measurement is solely from the noise, which encapsulates background noise and internal measurement noise. Thus, taking the mean of M0 measurements ð0;mÞ zi from sensor i gives an estimate of 2i ^2i ¼

M0 1 X ð0;mÞ z : M0 m¼1 i

ð22Þ

Similarly, if the source of unit acoustic power is placed at exactly unit distance from the sensor, then (21) reduces to ð1Þ

zi ¼ 2i þ i ;

ð23Þ

and so if M1 measurements are taken, then the sample mean gives an estimate of i ^2i þ î ¼

M1 X

1 ð1;mÞ z : M1 m¼1 i

c^ ¼

^2 Ns : 2^ z2

TABLE 3 Summary of c Estimated for Each Sensor for Different Distances

Maximum likelihood offers an alternative method for determining all of the parameters 2 , , and c, and in addition allows the calculation of . A set of samples is taken at varying distances from a source of known amplitude such that , which controls the decay of amplitude with distance in the measurement function, can be inferred, along with the other parameters if desired. Denoting sample m 2 f1 . . . Mg at distance d 2 D ¼ fd1 . . . dk g by zmd , the mean and variance of zmd are given by (9) and (10), respectively. The likelihood of all measurements z ¼ fzmd jm 2 f1 . . . Mg; d 2 Dg conditioned on the parameters ¼ f; ; 2 ; cg can then be written as pðzj Þ ¼ ¼

pðzmd j Þ

ð26Þ

2 2 : N zmd ; d þ 2 ; c d þ 2 Ns d2D m¼1 M YY

ð27Þ

The maximum likelihood solution for the parameters is calculated using ^ ¼ arg max log pðzj Þ

¼ arg min

M XX

ð28Þ

logðcÞ þ 2 log d þ 2

d2D m¼1

2 Ns zmd þ 1 : 2c d þ 2

ð29Þ

This is a nonlinear minimization problem which requires a numerical method to solve. Since 2 , , and c can be determined using the simple procedures outlined above, the minimization in (29) is simply over . Results of applying this calibration procedure for five different sensors gives the estimated parameters 2 , , and summarized in Table 2 and the parameter c summarized in Table 3 for each distance. From the estimates presented in the tables appropriate values of the constants and c, chosen to be independent of each mote, are

4.2

ð25Þ

M YY d2D m¼1

ð24Þ

Using (10) an estimate of c is possible by using the mean, ^ 2i , of a set of samples from a source of any zî , and variance, acoustic power. This gives

1337

2:2;

ð30Þ

c 1:5:

ð31Þ

Comparison of PF with a Simple Triangulation-Based Tracking System We first evaluate the performance of the PF-based tracking comparing it with a simple triangulation-based tracking system [2] for different number of nodes. For triangulation, we assumed that the distance estimates by measuring propagation delays are within þ= 5 percent of the actual

1338


VOL. 9,

NO. 9,

SEPTEMBER 2010

Fig. 1. Comparison of PF and triangulation-based tracking.

values. Simulation results for PF are obtained by using 5,000 particles and sampling interval of 1 sec. Fig. 1 shows RMS tracking error in the target’s location (with error bars representing the standard deviations). Note that there are no error bars for grid deployment for triangulation-based tracking. This is because the triangulation-based tracking always produces the same magnitude of error for a given grid topology (nonprobabilistic estimation). We used five different uniform random topologies and ran the simulation 100 times to calculate the standard deviations in the RMS values for both PF and triangulation. Fig. 1 shows that PF performs better than triangulationbased tracking for both grid and uniform random deployments. There are two main reasons for the poor performance by triangulation-based tracking. Firstly, the inaccuracies in distance estimates (taken as a random value between þ= 5 percent for this simulation) affect the triangulationbased localization and secondly for sparse topologies, the triangulation algorithm introduces errors when distance estimates are reported by less than three motes required for triangulation. The results show that PF performs much better than triangulation for sparse topologies and that difference between the performance of the two decreases for very dense topologies, e.g., 100 nodes in 120 m 120 m area.

2.53 meters for 49 nodes. For UR deployment, the corresponding value decreases from 6.96 to 3.11 meters. Grid deployment shows overall better performance than the UR deployment. Note that tracking performance for UR with 81 nodes is worse than that with 64 nodes. This can be attributed to the random nature of the UR deployment strategy. The UR deployment thus performs better or worse than the grid deployment depending on the node placement with respect to the target trajectory. Fig. 2 shows the computation time of the PF versus the number of deployed nodes. We can see that the computation time of the PF grows almost linearly with the number of deployed nodes for both grid and uniform random deployment. Packet loss also increases sharply with increase in number of nodes beyond 49 as shown in Fig. 3. This suggests, that for the given size of the deployment field, target and sensor node characteristics (intensity, sensitivity, etc.), and the selected networking and simulation parameters, the optimal density for the deployment of sensor nodes should be somewhere around 34 nodes per 100 m2 , which corresponds to 49 nodes for the target area under investigation. This node density gives a balance between the error in estimation (RMS tracking error of about 2.5-3 m) and total computation time (about 230 seconds).

4.3 Design Parameters We next evaluate the effect of various design parameters on the detection accuracy and computation time of the particlefilter-based tracking system. 4.3.1 Effect of the Number of Deployed Sensor Nodes Figs. 1a and 1b shows that for grid deployment, the RMS tracking error drops from about 7.14 meters for 16 nodes to

4.3.2 Effect of the Number of Generated Particles Next, we examine the effect of the number of generated particles on the performance of the tracking system. Recall that in our system, a number of particles are randomly generated near the sensor nodes. Further, the number of particles are independent of the number of deployed nodes (see Section 3) and only the Tpf component of the total computation time varies with change in number of particles.

Fig. 2. Effect of the number of nodes on the computation time.

Fig. 3. Effect of the number of nodes on packet loss.


1339

Fig. 4. Effect of the number of generated particles on the estimation error.

Fig. 4 shows the results of the simulation for different number of generated particles for both grid and UR deployments. The number of nodes is 49 and the sampling interval is 1 sec. In Fig. 4a, the RMS error in estimation for grid deployment reduces from about 4.28 to 2.37 meters when the number of particles are increased from 1,000 to 6,000. After that the error begins to decrease at a much slower rate. For UR deployment (Fig. 4b), the RMS tracking error reduces from 4.57 meters with 1,000 particles to 3.03 meters at 6,000 particles. However, from Fig. 5, we can see that for both grid and UR deployments the computation time of the PF grows linearly with the number of generated particles. Computation time for both grid and UR deployment is between 230 to 270 secs for 5,000-6,000 generated particles. This suggests that given the simulation parameters, 5,0006,000 particles is a good trade-off between detection accuracy and computation time.

4.3.3 Effect of the Sampling Interval Fig. 6 shows the effect of the variation in sampling interval on the PF’s behavior. The number of nodes is set as 49 with 5,000 particles. Intuitively, the larger the sampling interval, the longer the nodes will remain idle, hence the less energy they will consume and the longer the region can be surveilled. However, as can be seen from this figure, the RMS error in estimation of the target’s location begins to increase with the increase in the sampling interval. Best tracking results are obtained with 0.5 to 1 sec sampling interval for both grid and UR deployment. The computation time, on the other hand, reduces from about 500 to 230 sec as the sampling interval is increased from 0.5 to 1 sec in Fig. 7. This implies that given the node density and the selected

Fig. 5. Effect of the number of generated particles on the computation time.

simulation parameters, 1 sec sampling time is an appropriate choice for this case.

4.4 Effect of Calibration Parameters In Section 4.1, we have identified four calibration parameters that need to be determined for accurate tracking of a target. In this section, we evaluate the effect of variations in estimation of these calibration parameters on the performance of the particle-filter-based tracking system. For this set of simulations, we consider 49 nodes, 5,000 particles, and a sampling interval of 1 sec as recommended in the previous section. Also note that since changing the calibration parameters has no effect on the computation time, the results reported in this section does not include the computation time metric. 4.4.1 Effect of the Path Loss Factor of Target’s Intensity () Fig. 8 shows the RMS tracking error for different values of the path loss factor used in the PF for the target intensity. For this set of simulations, the synthetically produced target intensity values are based on path loss factor of 2.2 (30) while the path loss value supplied to the particle filter is varied from 1.6 to 3. The figure shows that if the value of the path loss factor (i.e., estimated through calibration) supplied to the PF does not correspond to the real value, RMS tracking error values increases sharply with the increase in the mismatch in value. Note that underestimation of the path loss factor (e.g., 1.6 instead of 2.2) results in higher error than its overestimation. 4.4.2 Effect of the Nonlinearity Constant (c) Fig. 9 shows the RMS tracking error for different values of the nonlinearity constant c used in the PF for the target intensity. The nonlinearity constant is varied from 1 to 2 while the synthetically produced target intensity values are based on the value of c as 1.5 (31). The figure shows that the mismatch between estimated and the real value of c has minimal effect on the tracking error for grid deployment. For UR deployment, the increase in error is also not steep with maximum observed tracking error of around 5 m. 4.4.3 Effect of the Gain Values () As described in Section 4.1, each sensor is calibrated to determine its individual gain values. For the synthetically produced target intensity values in simulations, random values of gain are assigned to each sensor (from a range of

1340


VOL. 9,

NO. 9,

SEPTEMBER 2010

Fig. 6. Effect of sampling interval on the estimation error.

valid values determined from calibration). In order to study the impact of variance in the gain values, we vary these values between þ=30 percent. Fig. 10 shows the RMS tracking error for different values of the gain () used in the PF for the target intensity. The RMS error values for both grid and UR deployment increase with increase/decrease in the gain values. Again the increase is not as high as observed for estimation of path loss (Section 4.4.1).

4.4.4 Effect of the Noise Variance Values (2 ) Similar to the gain values, each sensor is calibrated to determine its individual noise variance values (2 ). In order to study the effect of estimation of noise variance, we increased/decreased the randomly assigned noise variance values (selected from a valid range determined through actual calibration) for all the sensors from 5 to 20 percent. Fig. 11 shows that if the value of the noise variation (estimated through calibration) supplied to the PF does not correspond to the real value, RMS tracking error values increases sharply with the increase in the mismatch. Note that RMS error increases more when lower noise variance is estimated than the actual values as compared to the case when higher noise variance is estimated than the actual values. This is to be expected, as a conservative value for the measurement noise is less restrictive on the estimation process, whereas as an under-approximation of the measurement noise forces the estimation process to account for measurement errors in the estimate of the state. The RMS error for grid deployment increased from 2.53 to about 9 for a 10 percent decrease in the estimated values while the RMS error values become too high ð>20Þ

Fig. 7. Effect of sampling interval on the computation time.

for any further decrease after 10 percent. Increase in RMS tracking error is, once again, more for UR deployment as compared to grid deployment. The simulation results from Section 4.4 show that out of the four calibration parameters evaluated, estimation of noise variation values (2 ) has the most effect on the detection accuracy of the particle-filter-based tracking system. The tracking system can only tolerate about 10 percent lower estimated noise variance values. If the actual noise variance is more than 10 percent, the PF starts giving large tracking errors. The effect of estimation of noise variations is followed by the effect of path loss factor of target’s intensity () while estimation of gain () and nonlinearity constant (c) has relatively less effect on the performance of the tracking system. This highlights the need for an accurate calibration mechanism to estimate the critical values of noise variance and the path loss factor.

5

THEORETICAL ANALYSIS AND COMPARISON

In this section, we first derive the Crame´r-Rao Bound for the scenario used in our simulations. We next compare this theoretical lower bound with the results from our simulations.

5.1 The Posterior Crame´r-Rao Bound (PCRB) The Crame´r-Rao Bound (CRB) is a theoretical construct, which specifies a lower bound on the second-order estimation error performance of any unbiased estimator [5]. We compute this theoretical bound for the scenario used in the simulations of Section 4. The state evolution model and the measurement model are as described in Section 3. For simplicity, we will limit our analysis to the single target case, although the multitarget CRB directly follows [19]. As in the case of the simulation scenario, a WSN composed of N ¼ 49 acoustic sensor nodes is deployed in a two-dimensional region of size 120 m 120 m. An unbiased estimate of the state vector Xt , based on Z1:t and the known distribution of the initial target state, pðX0 Þ, ^ t , and its error covariance matrix by Pt . The is denoted by X lower bound of Pt , referred to as the posterior Crame´r-Rao bound is expressed as follows [5]: n o : ^ t Xt ÞðX ^ t Xt ÞT jz1:t ; pðX0 Þ J 1 : ð32Þ Pt ¼ E ð X t


1341

Fig. 8. Effect of path loss factor on the estimation error.

Fig. 9. Effect of nonlinearity constant on the estimation error.

Fig. 10. Effect of gain on the estimation error.

Fig. 11. Effect of noise variance on the estimation error.

The inequality in (32) implies that the difference Pt Jt1 is a positive semidefinite matrix. Matrix Jt in (32) is referred to as the information matrix and its inverse is the PCRB. For a nonlinear filtering problem specified by (1) and (11), the information matrix can be computed recursively as follows [1]: tþ1 ð33Þ Jtþ1 ¼ ½F 1 T Jt F 1 E X Xtþ1 log pðZtþ1 jXtþ1 Þ ;

where r ¼

@ @ ;...; @1 @r

T ¼ r r ;

T ;

ð34Þ ð35Þ

and the likelihood function, pðZtþ1 jXtþ1 Þ, is defined by (9) and (10). For the purposes of calculating the bound it is assumed that the process noise, Q, is zero, that is the target

1342


VOL. 9,

NO. 9,

SEPTEMBER 2010

Fig. 13. Experimental setup. Fig. 12. Comparison of CRB and simulation results.

trajectory is deterministic. For the measurement model specified in Section 3.2, the expectation in (33) can be evaluated as N X tþ1 T Htþ1;i R1 E X tþ1;i Htþ1;i ; Xtþ1 log pðZtþ1 jXtþ1 Þ ¼

ð36Þ

i¼1

where Ht;i is the Jacobian of hi ðXt Þ evaluated at the true value of Xt 2 3 ðxi xt Þi It d2 t;i 6 7 0 6 7 2 7 ð37Þ Ht;i ¼ 6 ðy y Þ I d i t i t t;i 6 7; 4 5 0 i d t;i and R1 t;i is given by R1 t;i ¼

4c þ Ns 2 ; 2 2c i þ i It d t;i

ð38Þ

once again evaluated at the true value of Xt . The terms in (37) and (38) are as defined in Section 3.

5.2 Comparison of PCRB and Simulation Results For the PCRB calculations, the initial state vector was chosen as X0 ¼ ½15

0:25

5 0:25

0:7T ;

ð39Þ

where the first and the third components are in meters, while the second and the fourth components are in meters/sec. For the purposes of this work, the initial J0 was approximated as a diagonal matrix with J01 ¼ P0 ¼ diagð½0:1 0:05 0:1 0:05

0:012 Þ:

ð40Þ

The displayed error bounds are computed as follows: qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð41Þ bt ¼ Jt1 ½1; 1 þ Jt1 ½3; 3; where Jt1 ½1; 1 and Jt1 ½3; 3 are the diagonal elements of the inverse of the information matrix corresponding to the x and y coordinates, respectively. Next, we compare the theoretical lower error bounds with the RMS errors of the simulated PF. The number of particles used was 5,000 and the PF was run 500 times for the bound comparison. Fig. 12 illustrates that the simulated PF follows the general trend of the theoretical bound. We

expect some mismatch between the filter performance and the bound due to the fact that the bound is a theoretically ideal quantity.

6

EXPERIMENTS

In order to validate the simulations results described in Section 4 and to test the performance of the PF-based tracking system in a real-world environment, we have conducted initial experiments. In this section, we present our findings obtained from these indoor experiments. Fig. 13 shows the setup of our test bed. We used a prototype test bed consisting of Xbow MicaZ motes. These motes were programmed to perform high frequency sampling to measure the acoustic signals (at 5 kHz) generated by the target (a remote controlled toy car). We used three different topologies where 8, 10, and 14 of these MicaZ motes were deployed in an area of size 2:0 meters 3:5 meters (Fig. 13) with different internode spacing. A mote attached with a laptop was used as the data sink. The motes measure the acoustic signals and send a summary of the acoustic samples to the laptop (single hop communication), which executes the PF. The speed of the target was varied from 0.2 to 0.35 m/ sec in different set of experiments. Time synchronization among the motes was achieved by a Beeper mote broadcasting a beacon message every 0.5 seconds (sampling interval). On receiving a beacon, each mote starts sampling to collect 1,000 discrete samples (taking 0.2 s at 5 kHz). To reduce the storage requirements at the motes, radio communications, and the size of the data packets, motes only maintain summary statistics (sum and sum of squares) instead of the raw values. These summary statistics were then transferred during the next 0.3 seconds over the wireless link to the base station. Note that the sample variance can be easily calculated at the base station as we know the total number of samples, sum, and sum of squares of the samples, as shown !2 Ns Ns 1X 1X 2 ðXi Þ Xi : ð42Þ varðXÞ ¼ N i¼1 N i¼1 To ensure that all data packets are received at the base station, we incorporated an application level TDMA scheme to allocate discrete contention-free time slots to each mote for data transfer to the base station (discussed in detail in Section 7). The base station executes the PF with 5,000 particles. Prior to running the experiments, the environmental path loss factor and nonlinearity constant are


1343

Fig. 14. Tracking results (arrows represent actual paths).

estimated through calibration experiments. Moreover, gain and noise variation are also estimated by calibrating each mote carefully in the same experimental environment, as described in Section 4.1. Some of the tracking results are shown in Fig. 14 where actual paths are shown by pointed arrows. Fig. 15 shows the quantitative performance (mean and standard deviation of tracking error) of the WSN-based tracking system for two different velocities of the target. Results show that mean tracking error is less for cases when more motes are deployed and when speed of the target is slower. Best tracking results (mean error of 0.11 m) is obtained for 14 motes in the topology with speed of target at 0.2 m/s. Note that the reported errors are approximate in that they have been calculated at discrete points along the path due to variability in speed of the target. The PF implemented in Matlab (running on a 1.6 GHz laptop with 1 GB of RAM) takes on average about 200 ms to estimate the next target position for one iteration of measurements. This shows that the PF can be implemented in an online tracking system to give close to real-time tracking results. The results demonstrate that the PF-based estimator performs very well in tracking of the target using real data collected by off-the-shelf WSN devices.

7

SYSTEM DESIGN CHALLENGES

In this section, we discuss various system design challenges faced during the experimental study. We also describe the steps taken to overcome these challenges. We believe that these contributions are particularly important for realworld field deployment of WSN systems.

7.1 Calibration Our empirical experiments have shown that estimation of correct calibration parameters have a significant effect on the performance of the PF-based estimator. Initially, we started with a simple calibration mechanism where sensors were placed at known distances from a target acoustic signal (e.g., a 4 kHz tone emitted by a sounder of MicaZ acoustic sensor board) of known intensity. These sensors are then characterized in term of mean and variance of received total acoustic power as a function of the distance. The same set of calibration experiments is then repeated with no sound source to estimate the background noise. These intensity and background noise values are then fed into the PF for estimation. The results obtained by applying this calibration procedure to the tracking system indicated that the calibration mechanism need to be modified for improving the performance.

Fig. 15. Quantitative tracking results.

An improved version of the calibration process was later adopted where calibration is done with a white noise source of known signal power (Section 4.1). The improved calibration process is simpler and has shown better results than its predecessor process.

7.2 Reliability Issues Packet losses are expected to occur in a real multihop network deployment. These losses can be due to channel contention, collisions, and queues overflows in buffers. To determine the maximum packet transmission rate in order to avoid the buffer overflows, a source mote was initially programmed to send 34 bytes packets to a mote base station running the standard TinyOS TOSBase application. The packet transmission rate was progressively increased until the base station started dropping packets at its radio/UART queues. The base station was able to receive all of the packets (about 128 packets/sec) with interpacket transmission delay of 8 ms. This gives us a bound on the maximum transmission rate attainable with the default setup. To avoid the packet losses due to contention and collisions, we incorporated an application level TDMA in our implementation with the time slot defined as 8 ms (for multiple senders-single receiver). Recall that motes collect 1,000 samples in 0.2 sec (@5 kHz) and have the next 0.3 seconds to transmit a 34 byte packet containing the summary statistics of these 1,000 samples. With 8 ms as the interpacket transmission delay to avoid the buffer overflows, we can have 34 nodes in the collision domain transmitting in TDMA time slots giving us 100 percent successful packet transmission rate. Note that we have used a simple application layer TDMA scheme instead of using a MAC layer TDMA [20] in order to match the application sampling rate with the TDMA interval. 7.3 Scalability Issues In order to make the implementation scalable, we introduced clustering based on multiple collision domains by tuning motes to different transmission channels (16 available channels in Xbow 2.4 GHz MicaZ motes). These clusters form the lower layer of a hierarchical topology. This clustering introduces additional issues such as time synchronization among different collision domain-based clusters and data fusion among multiple base stations. For time synchronization, we chose to synchronize the beeper nodes with a hybrid time synchronization protocol that

1344


incorporates accurate byte level time stamping from Flooding Time Synchronization Protocol (from Vanderbilt University) [21] in simple pairwise message exchanges of the Time synchronization Protocol for Sensor Networks (TPSN) protocol [22]. The time stamps are made at each byte boundary after the SYNC bytes as they are transmitted or received to counter the nondeterministic delays in radio message delivery in WSN. We have conducted further experiments with multiple collision domains with 25 nodes [17] and 36 nodes grid topologies (not reported in this paper due to space limitations). All nodes were able to successfully transmit their data packets to their respective base stations.

[2]

[3]

[4] [5] [6]

[7]

8

CONCLUSION AND FUTURE WORK

In this paper, we have presented a detailed evaluation of the performance of a WSN tracking system, which employs the PF algorithm. We developed a simple yet effective mechanism for calibrating off-the-shelf sensor motes. We observed the effect of various design and calibration parameters such as number of deployed nodes, number of generated particles, sampling interval, path loss factor, noise variations, gain, and nonlinearity constant on the estimation accuracy and computation time of the tracking system. We also made recommendations on the optimal values for these parameters, which achieve an acceptable trade-off between accuracy and complexity. Moreover, we identified that among the calibration parameters, estimation of the noise variation values has the most significant effect on the performance of the tracking system. It is pertinent to mention here that the characteristics of the target such as its volume and speed also effect the choice of these design parameters. These recommendations on the choice of design parameters can be utilized to engineer a real WSNbased tracking solution. We also derived the theoretical lower bound, the CRB, and demonstrated that the results from our simulations are comparable to the bound. We discuss the challenges experienced in implementing a prototype of our system using off-the-shelf sensor nodes. We describe custom algorithms that were implemented for successfully realizing the system. We found that our prototype comprising of off-the-shelf sensor nodes can successfully track the moving target in several scenarios. Our experimental results are quite promising and suggest that low-cost sensor nodes can effectively complement sophisticated target detection hardware. In future, we plan to do a real-time implementation of the PF-base tracking system and to test the implementation in the outdoor environment.

[8] [9]

[10]

[11]

[12] [13] [14] [15]

[16] [17]

[18] [19] [20]

[21]

ACKNOWLEDGMENTS This paper is a result of collaboration between UNSW (School of Computer Science and Engineering) and DSTO (ISR Division).

REFERENCES [1]

B. Ristic, S. Arulampalam, and N. Gordon, Beyond the Kalman Filter: Particle Filters for Tracking Applications. Artech House, 2004.

[22]

VOL. 9,

NO. 9,

SEPTEMBER 2010

Q. Wang, W. Chen, R. Zheng, K. Lee, and L. Sha, “Acoustic Target Tracking Using Tiny Wireless Sensor Devices,” Proc. Second Int’l Conf. Information Processing in Sensor Networks (IPSN ’03), pp. 642657, 2003. L. Gu, D. Jia, P. Vicaire, T. Yan, L. Luo, A. Tirumala, Q. Cao, T. He, J.A. Stankovic, T. Abdelzaher, and B.H. Krogh, “Lightweight Detection and Classification for Wireless Sensor Networks in Realistic Environments,” Proc. ACM SenSys, pp. 205-217, 2005. M.F. Duarte and Y.H. Hu, “Vehicle Classification in Distributed Sensor Networks,” J. Parallel and Distributed Computing, vol. 64, no. 7, pp. 826-838, 2004. H.L. Van Trees, Detection, Estimation and Modulation Theory (Part I). John Wiley and Sons, 1968. A. Ledeczi, A. Nadas, P. Volgyesi, G. Balogh, B. Kusy, J. Sallai, G. Pap, S. Dora, K. Molnar, M. Maroti, and G. Simon, “Countersniper System for Urban Warfare,” ACM Trans. Sensor Networks, vol. 1, no. 2, pp. 153-177, 2005. T. He, S. Krishnamurthy, J.A. Stankovic, T.F. Abdelzaher, L. Luo, R. Stoleru, T. Yan, L. Gu, J. Hui, and B. Krogh, “An EnergyEfficient Surveillance System Using Wireless Sensor Networks,” Proc. ACM MobiSys, pp. 270-283, 2004. M.J. Coates and G. Ing, “Sensor Network Particle Filters: Motes as Particles,” Proc. IEEE Workshop Statistical Signal Processing, pp. 1152-1157, 2005. C. Taylor, A. Rahimi, J. Bachrach, H. Shrobe, and A. Grue, “Simultaneous Localization, Calibration, and Tracking in an Ad Hoc Sensor Network,” Proc. Fifth Int’l Conf. Information Processing in Sensor Networks (IPSN ’06), pp. 27-33, 2006. B. Kusy, J. Sallai, G. Balogh, A. Ledeczi, V. Protopopescu, J. Tolliver, F. DeNap, and M. Parang, “Radio Interferometric Tracking of Mobile Wireless Nodes,” Proc. Fifth Int’l Conf. Mobile Systems, Applications and Services (MobiSys ’07), pp. 139-151, 2007. L. Klingbeil and T. Wark, “A Wireless Sensor Network for RealTime Indoor Localization and Motion Monitoring,” Proc. Seventh Int’l Conf. Information Processing in Sensor Networks (IPSN ’08), pp. 39-50, 2008. M. Bolic, P.M. Djuri, and S. Hong, “Resampling Algorithms and Architectures for Distributed Particle Filters,” IEEE Trans. Signal Processing, vol. 53, no. 7, pp. 2442-2450, July 2005. R. Brooks, P. Ramanathan, and A. Sayeed, “Distributed Target Classification and Tracking in Sensor Networks,” Proc. IEEE, vol. 91, no. 8, pp. 1163-1171, Aug. 2003. D.J. Salmond and H. Birch, “A Particle Filter for Track-BeforeDetect,” Proc. Am. Control Conf., 2001. T. Schon, F. Gustafsson, and P.-J. Nordlund, “Marginalized Particle Filters for Mixed Linear/Nonlinear State-Space Models,” IEEE Trans. Signal Processing, vol. 53, no. 7, pp. 2279-2289, July 2005. M.G. Rutten, N.J. Gordon, and S. Maskell, “Recursive Trackbefore-Detect with Target Amplitude Fluctuations,” IEE Proc. Radar Sonar Navigation, vol. 152, no. 5, pp. 345-352, Oct. 2005. N. Ahmed, M. Rutten, T. Bessell, Y. Dong, S. Kanhere, N. Gordon, and S. Jha, “Performance Evaluation of a Wireless Sensor Network Based Tracking System,” Proc. Fifth IEEE Int’l Conf. Mobile Ad Hoc and Sensor Systems (MASS ’08), Oct. 2008. K. Whitehouse and D. Culler, “Calibration as Parameter Estimation in Sensor Networks,” Proc. ACM Int’l Workshop Wireless Sensor Networks and Applications (WSNA ’02), pp. 59-67, 2002. B. Ristic and M. Morelande, “Cramer-Rao Bound for Multiple Target Tracking Using Intensity Measurements,” Proc. Conf. Information, Decision and Control (IDC ’07), pp. 354-259, Feb. 2007. K.K. Chintalapudi and L. Venkatraman, “On the Design of Mac Protocols for Low-Latency Hard Real-Time Discrete Control Applications over 802.15.4 Hardware,” Proc. Int’l Conf. Information Processing in Sensor Networks (IPSN ’08), pp. 356-367, Apr. 2008. M. Maroti, B. Kusy, G. Simon, and A. Ledeczi, “The Flooding Time Synchronization Protocol,” Proc. Second ACM Conf. Embedded Networked Sensor Systems (SenSys ’04), pp. 39-49, Nov. 2004. S. Ganeriwal, R. Kumar, and M.B. Srivastava, “Timing-Sync Protocol for Sensor Networks,” Proc. First ACM Conf. Embedded Networked Sensor Systems (SenSys ’03), pp. 138-149, Nov. 2003.


Nadeem Ahmed received the BE degree from the University of Engineering and Technology, Lahore, Pakistan, and the MS and PhD degrees in computer sciences from the University of New South Wales (UNSW), Sydney, Australia, in 2000 and 2007, respectively. He is currently a research associate with the School of Computer Science and Engineering at UNSW. His research interests include wireless sensor networks and mobile ad hoc networks. He is a member of the IEEE. Mark Rutten received the BSc degree in computer science and theoretical physics in 1994, the BE degree in electrical engineering in 1995, and the MSc degree in signal and information processing in 1998, all from The University of Adelaide, and the PhD degree in electrical and electronic engineering from the University of Melbourne in 2005. Since 1996, he has been working with the Defence Science and Technology Organization in Edinburgh, Australia, working with the Tracking and Sensor Fusion group from 1999. His research interests include nonlinear filtering, tracking, and track fusion. Travis Bessell received the BE degree in computer systems engineering from Flinders University in 2005 and the MSc degree in signal and information processing from The University of Adelaide in 2008. Since 2005, he has been working with the Defence Science and Technology Organization, Edinburgh, Australia, working with the Tracking and Sensor Fusion group. His research interests include wireless sensor networks, nonlinear filtering, and tracking.

Salil S. Kanhere received the BE degree in electrical engineering from the University of Bombay, India, in 1998, and the MS and PhD degrees in electrical engineering from Drexel University, Philadelphia, in 2001 and 2003, respectively. He is currently a senior lecturer with the School of Computer Science and Engineering at the University of New South Wales, Sydney, Australia. His current research interests include wireless sensor networks, vehicular communication, mobile computing, and network security. He is a member of the IEEE and the ACM.

1345

Neil Gordon received the BSc degree in mathematics and physics from Nottingham University, United Kingdom, in 1988, and the PhD degree in statistics from Imperial College, University of London, in 1993. From 1988 to 2002, he was with various research groups within DERA and QinetiQ working in the areas of missile guidance, target tracking, and statistical data processing. Since August 2002, he has been with DSTO in Australia where he is currently the head of the Tracking and Sensor Fusion research group. He is a coeditor/coauthor of two books on particle filtering. He is a member of the IEEE. Sanjay Jha received the PhD degree from the University of Technology, Sydney, Australia. He is a professor and head of the Network Group in the School of Computer Science and Engineering at the University of New South Wales. His research activities cover a wide range of topics in networking including wireless sensor networks, ad hoc/community wireless networks, resilience/quality of service (QoS) in IP networks, and active/programmable network. He has published more than 100 articles in high quality journals and conferences. He is the principal author of the book Engineering Internet QoS and a coeditor of the book Wireless Sensor Networks: A Systems Perspective. He is an associate editor of the IEEE Transactions on Mobile Computing. He was a member-at-large for the Technical Committee on Computer Communications (TCCC) of the IEEE Computer Society for a number of years. He has served on program committees of several conferences. He was the technical program committee chair of the IEEE Local Computer Networks LCN2004 and ATNAC04 conferences, and the cochair or general chair of the Emnets-1 and Emnets-II workshops, respectively. He was also the general chair of the ACM SenSys 2007 symposium. He is a senior member of the IEEE and the IEEE Computer Society.

. For more information on this or any other computing topic, please visit our Digital Library at www.computer.org/publications/dlib.