Collaborative Sensing to Improve Information Quality for Target Tracking in Wireless Sensor Networks

Wendong Xiao and Chen Khong Tham Institute for Infocomm Research

Agency for Science, Technology and Research (A * Star) Singapore

{wxiao, cktham}@i2r.a-star.edu.sg

Abstract- Due to limited network resources for sensing, communication and computation, information quality (IQ) in a

wireless sensor network (WSN) depends on the algorithms and protocols for managing such resources. In this paper, for target tracking application in WSNs consisting of active sensors (such as ultrasonic sensors) in which normally a sensor

senses the environment actively by emitting energy and measuring the reflected energy, we present a novel collaborative sensing scheme to improve the IQ using joint sensing and adaptive sensor scheduling. With multiple sensors

participating in a single sensing operation initiated by an emitting sensor, joint sensing can increase the sensing region of an individual emitting sensor and generate multiple sensor measurements simultaneously. By adaptive sensor scheduling,

the emitting sensor for the next time step can be selected adaptively according to the predicted target location and the detection probability of the emitting sensor. Extended Kalman

filter (EKF) is employed to estimate the target state (i.e., the

target location and velocity) using sensor measurements and to predict the target location. A Monte Carlo method is presented to calculate the detection probability of an emitting sensor. It is demonstrated by simulation experiments that collaborative

sensing can significantly improve the IQ, and hence the tracking accuracy, as compared to individual sensing.

Keywords-information quality; wireless sensor networks; target tracking; joint sensing; sensor scheduling; collaborative sensing; Kalman jiber

I. INTRODUCTION

Typically, a wireless sensor network (WSN) is application-driven and mission-critical. Therefore, the information quality (lQ) or quality of information (QoI) such as the accuracy of target tracking or event detection is critical for the end users, service providers and the system designers. To provide accurate IQ in WSNs is challenging due to the resource-constrained, dynamic and distributed nature of the network and the lack of a holistic design approach, which takes into account different types of resources and their inter-dependencies.

Recently, IQ is receiving increasing interests for various WSN applications. For example, in [1], based on dynamic Bayesian network, sensor selection approaches for human activity detection are proposed to optimize IQ represented by the entropy of the detection probability. In [2], the

978-1-4244-5328-3/10/$26.00 ©2010 IEEE 99

Sajal K. Das Department of Computer Science and Engineering

The University of Texas at Arlington Arlington, TX, USA

das@uta.edu

relationship between the sensor sampling rate and the QoI metric of timeliness and confidence is derived. In [3], the entropy of sensory data is used to quantify the IQ. Based on the exponential correlation model for the sensory data, an asynchronous sampling strategy is proposed to improve the IQ through shifting the sampling moments of sensors.

Target tracking in WSNs has been studied extensively. Due to the limited sensing capability and limited resources for communications and computation, collaborative resource management is required to trade-off between the tracking accuracy, i.e., the IQ of the target tracking application, and the resource usages, e.g., through selection of single tasking sensor [4-6] or multiple tasking sensors [7]. A distributed market-based congestion scheme is presented in [8] for competition of allocated time slot in a node among multiple target tracks with different QoI and priorities. Ultrasonic WSN test-beds for target tracking are also developed using centralized architecture [9] and distributed sensor competition [10] to show the IQ of different sensor scheduling schemes.

In general, sensors used in WSNs can be classified into active and passive ones. Passive sensing mechanism is used in acoustic, seismic or thermal sensors where the sensor measures energy already in the environment [5, 7]. A sensor adopting the active sensing mechanism, like the ultrasonic sensor, senses the environment actively by emitting energy and measuring the reflected energy [6, 9, 10].

To the best of our knowledge, in the existing literature, the tasking sensor works independently of other sensors for individual measurement. In this paper, based on active ultrasonic sensors, we will introduce a joint sensing mechanism by using single sensor to emit the energy and multiple sensors to measure the reflected energy signals from the target, and present a novel collaborative sensing scheme using joint sensing and adaptive sensor scheduling to select the emitting sensor for the next time step according to the predicted target detection probability of the emitting sensor. We will show that joint sensing can increase the sensing region of the emitting sensor and enable more useful sensor measurements simultaneously and the collaborative sensing can improve the IQ significantly.

Due to the nonlinear characteristic of the measurements of joint sensing, extended Kalman filter (EKF) [11] will be

Authorized licensed use limited to: Sethu Institute of Technology. Downloaded on July 08,2010 at 05:21:03 UTC from IEEE Xplore. Restrictions apply.

used to predict and estimate the target location and velocity information. The predicted target location will then be used to obtain the predicted target detection probability based on its Monte Carlo sampling.

The paper is organized as follows: Joint sensing mechanism will be introduced in Section II. EKF for fusing sensor measurement and calculating IQ for target tracking will be described in Section III. The collaborative sensing scheme, including the adaptive sensor scheduling, target detection model, target detection probability and its corresponding Monte Carlo calculate method, will be

detailed in Section IV. Simulation results will be reported in Section V. Finally conclusions and future work will be introduced in Section VI.

II. JOINT SENSING

In this paper, we assume that each ultrasonic sensor installs the sound wave emitter and receiver, and all the sensors in the network are homogeneous and time synchronized.

Normally an ultrasonic sensor adopts the active sensing mechanism where the sensor emits sound wave and measures the reflected echo from the target. The time of flight (TOF) is converted into range information towards the target. In this paper, we adopt a simplified cone shape detection region model for a typical ultrasonic sensor, where one ultrasonic sensor i is characterized by its location (xs;,

YSi), orientation fl;, detection angle a , and detection range

d. The TOF equals to the round trip time of the wave from the emitting sensor to the target and then back to the emitting sensor, which corresponds to the round trip distance of the sound wave that is bounded by 2d.

As shown in Fig. 1, only when the target is within the detection region (e.g., at location A) of emitting ultrasonic sensor 3, this sensor can obtain its measurement individually. Sensor 3 can never detect the target when the target is outside its detection region (e.g., at location B), although the sound wave can also reach ultrasonic sensor 5 after being reflected from the target (because the total trip distance is less than 2<1). This signal received by ultrasonic sensor 5 is simply discarded in [5, 8]. However, we found that such signal can be very useful as the sensor measurement, therefore we call ultrasonic sensor 5 can jointly sense the target with the emitting sensor 3. Fig. 1 also shows that when the target is located (e.g., at location A) in the detection region of the emitting sensor 3, joint sensing can also be done by sensor 2. Note that sensor 6 can not jointly sense any target with sensor 3 as the sound wave from ultrasonic sensor 3 can always reach ultrasonic sensor 6 directly, no matter whether or not there are targets in the network.

In this paper, we assume that the target can be jointly sensed by two sensors, if the following joint sensing conditions are satisfied: 1. The target is within the detection angles of both sensors; 2. The sum of distances from the target to the sensors is

less than 2d;

100

3. The two sensors are not within line of sight with each other (i.e., not within the detection angle of each other).

According to the above joint sensing conditions, no matter which sensor in these two sensors is the emitter, the signal can be received by the other sensor.

As an example, Fig. 2 shows the joint sensing region of sensors 1 and 2, when sensor 1 is the emitting sensor and sensor 2 is the receiving sensor. The ellipse consists of all points where the sum of its distances to sensor 1 and sensor 2 is 2d. The target must be inside this ellipse if sensor 1 and sensor 2 can jointly sense the target.

Ultrasonic

150

100

50

Ultrasonic Ultrasonic

Sensor5C'ollabonlth·e Sensor 4

Ultrasonic �/b===�� Sensor3

°OL--50�- 1�00---150�����--�=-���-450L--500� Ultrasonic Sensor 1 Sensor 2

Figure l. Joint sensing

Figure 2. Joint sensing region

Figure 3. Joint sensing region of sensors 1, 2 and 3

Authorized licensed use limited to: Sethu Institute of Technology. Downloaded on July 08,2010 at 05:21:03 UTC from IEEE Xplore. Restrictions apply.

In Fig. 2, areas a and b can only be sensed by sensor I individually, as any point in area a is not in the detection angle of sensor 2 (i.e., not satisfy joint sensing condition 1) and the sum of the distances from any point in area b to sensor 1 and sensor 2 is larger than 2d (i.e., not satisfy joint sensing condition 2). Areas c, d, and e can be jointly sensed by sensor 1 and sensor 2 as any point in them satisfies the three joint sensing conditions.

Similarly, we know that areas f and g can not be jointly sensed by sensor 1 and sensor 2.

The original detection region of sensor 1 by individual sensing is the union of areas a, b, c and d. Now we fmd that by joint sensing, the target located in area e can also be jointly sensed, which indicates that joint sensing can increase the detection region of individual sensors. In addition, if the target is located in area c or d, we can obtain two sensor measurements, one is the distance from sensor 1 to the target, and the other one is the sum of the distances from sensor 1 to the target and from the target to sensor 2.

As another example, Fig. 3 shows the joint sensing region of sensors 1, 2 and 3 when sensor 1 is the emitting sensor. We can fmd that the detection region of the joint sensing is irregular and much larger than the original individual sensor detection region.

III. EKF TRACKING ALGORITHM AND IQ

The EKF algorithm is adopted to fuse the joint sensing measurements and the measurements taken at different time steps.

The following constant velocity target motion model is used in this paper:

X(k+l) = F.0(k) +qU(k) ( 1)

with

X(k) = x,,(k+ 1) R

[�k+l) 1 y(k+l)

, k

Yik+l)

and

1 Nk 0 0 o 1 0 0 0 0 1 Nk 0 0 o 1

U(k)=(Ux(k») uik)

Mi 0 2

Mk 0 G(k)=

Mi 0 2

0 Mk

where x(k) and y(k) are x- and y- coordinates of the target at

time step k; xv(k) and yv(k) are respectively the

velocities of the target along x- and y- directions at time step

k; At k is the time difference between the measurement

times at steps k and k+ 1. Here U(k) is the Gaussian white acceleration noise with zero mean and covariance matrix Q.

Suppose at time step k, sensor s(k) is the emitting sensor, and the target is jointly sensed by s(k) and sensors sl(k), S2(k), . . . , sm(k), then the nonlinear observation model is

101

Z(k)= hk(X(k),Xs(k),Xs (k),Xs (k), ... ,Xs (k»+V(k) 1 2 m

with

h(X(k), Xs (k),Xst (k»

h(X(k),Xs(k),Xs (k» = .

2 +V(k)

h(X(k), Xs (k),Xsm (k»

(2)

J(X(k),X.(k),X.j (k»=�(x(k)-Xs(k/ +(y(k)-YS(k/ +�(x(k)-XSj(k/ +(y(k)-Y�(k/ (3)

where Xs(k) = (Xs(k),Ys(k)) is the known location of sensor

s(k), andX:j (k)=(x� (k�y� (k)) is the location of the sensor

Sj(k). Here V (k) is the Gaussian white measurement noise

of the receiving sensors with zero mean and variance R(k).

To be consistent, when the target can be sensed by emitting sensor s(k), the measurement equation (3) is also applicable with the emitting sensor and the receiving sensor being the same.

Given the estimate X(k I k) of X(k) at time step k,

EKF algorithm consists of an prediction update phase to

calculate the predicted state X (k + 11 k) and its

corresponding prediction error covariance P( k + 1 I k) , and a measurement update phase to obtain the new state

estimation X (k + 11 k + 1) and its corresponding error

covariance matrix �k+llk+l) from X(k + llk) and

P(k + 11 k) using the sensor measurements Z(k+ 1). The detailed information on EKF operations can be

found in [ 1 1] and is omitted here due to the space limitation.

Based on the state estimation, various measures can be defmed for the IQ, i.e., the tracking accuracy, such as the trace or the determinant of the covariance matrix, eigenvalues of the difference between the desired and the predicted covariance matrices, and the entropy of the state

estimation distribution. In this paper, the IQ, <p( k) , at time

step k is defined as the trace of the covariance matrix, i.e.,

¢(k) = Trace(P(k I k)). (4)

IV. COLLABORATIVE SENSING AND ADAPTIVE SENSOR

SCHEDULING

A. Inter-Sensor Interference and Sensor Scheduling

A serious problem in WSN of active sensors is the inter-sensor interference (lSI) when nearby ultrasonic

Authorized licensed use limited to: Sethu Institute of Technology. Downloaded on July 08,2010 at 05:21:03 UTC from IEEE Xplore. Restrictions apply.

sensors emit sound wave simultaneously. Such interference will result in erroneous sensor readings and must be dealt with properly. lSI also introduces a new technological constraint in the design and implementation of a WSN. In this paper, we assume the WSN is deployed in a small area where the sensor nodes are in the interference range of each other, and only single target tracking is considered. To avoid lSI, at each time step, only one emitting sensor will be scheduled and the other sensors will participate in joint sensing with the scheduled emitting sensor.

Periodic sensor scheduling is used in [6, 9] where the time is divided into periodic cycles. Within a cycle, a predefmed duration (called time slot) is assigned for each ultrasonic sensor for sensing, during which it can work properly without interference from other sensors.

A critical drawback of periodic sensor scheduling is that a detection may be missed when a scheduled sensor can not generate effective joint sensing measurements, which results in lower tracking accuracy. This problem is shown in Fig. 4 for a WSN with six ultrasonic sensors. The discretized target trajectory is displayed as circles with the scheduled sensors indicated beside the trajectory points, with the trajectory point displayed as shaded circle if it can be successfully detected by joint sensing of sensors, otherwise displayed as non-shaded circles. In the scheduling cycle i identified by the solid ellipse, the sensors are scheduled from sensor 1 to sensor 6. However, only scheduled sensors 1, 3,

5, and 6 generate effective detections whereas sensors 2 and 4 generate empty detections. For example, the first trajectory point with scheduled emitting sensor 1 associates with two measurements by sensors 1 and 6. The second trajectory point with scheduled emitting sensor 2 can not be detected as it is outside the detection angle of sensor 2. The third trajectory point can be detected by joint sensing of sensor 3 and sensor

5. Similarly, in the scheduling cycle i+ 1, only scheduled sensors 2 and 6 generate effective detections whereas sensors 1, 3, 4, and

5 generate empty detections.

To overcome the above drawback of period sensor scheduling, we introduce the adaptive sensor scheduling to select the emitting sensor for the next time step according to the predicted target location and the sensing region of the sensors.

Figure 4. Effective and missing detections in periodic sensor scheduling

102

B. Collaborative Sensing

In this paper, collaborative sensing is used to stand for joint sensing and the joint sensing enabled adaptive sensor scheduling.

Either centralized or distributed target tracking structure can be adopted, depending on the fusion centre being the centralized management centre or the scheduled sensor. At each time step, the scheduled sensor emits the sound wave and all other sensor nodes that can perform joint sensing with the emitting sensor will collect the measurements and forward the measurements to the fusion centre. The fusion centre will run EKF to give updates of the state estimation using the new measurements and schedule the emitting sensor for the next time step. Then it will inform the scheduled sensor to perform the emitting operation in the next time step, together with the state estimation and covariance matrix information in the distributed structure. We assume that the fusion centre knows the location and orientation of each sensor. In the distributed structure, this means that each sensor knows such information of each node because each sensor is possible to be the fusion centre.

Different measures can be used as the performance indices to select the emitting sensor, including the joint sensing detection probability, tracking accuracy, and energy efficiency. However, to calculate these performance indices under the joint sensing mechanism is not an easy task. For simplification and easy to compare with individual sensing scheme, in this paper, we schedule the emitting sensor according to the individual sensor detection probability.

Due to the uncertainties in the target motion model such as the target maneuvering, even using adaptive sensor scheduling, it is still possible that the scheduled sensor can not detect the target. If this happens, the fusion centre will use the predicted state and its covariance matrix as the estimation result.

C. Detection Probability for Individual Sensing

Suppose Sj is the emitting sensor, for a given target location X=(x, y), the target can be detected by Sj individually if it is in the detection region of Sj. Mathematically the following target detection model is used

p. (x,y)

= {I, if (x,y) is in the det�tion regionof Sj (5)'Ij; I 0, otherwtse.

Without loss of generality, in this paper, we suppose the target location is 2-dimensional. For emitting sensor selection, the prediction of the detection probability of a given emitting sensor is required. Denote the prediction of

the target location and its covariance as J.1 and L which

are sub-vector of X(k + 11 k) and sub-matrix of P(k+llk)

respectively. Then the probability density function (PDF) function of the 10cationX=(x, y) will be

f(x,y) = ��112 ext{ �( X -J.1t r;1 (X -J.1) ). (6)

Authorized licensed use limited to: Sethu Institute of Technology. Downloaded on July 08,2010 at 05:21:03 UTC from IEEE Xplore. Restrictions apply.

State prediction for step k+1

State estimation at step k

Sensor Sj

Figure 5. Predicted detection probability for joint sensing

State prediction for step k+1

Sensor Sj

Figure 6. Approximate the Gaussian distribution by random samples

Fig. 5 shows the current state estimation and state

prediction for the next time step by 3a ellipses, as well as

the sensing region of sensor Sj. In general, the prediction of the joint sensing detection probability by using Sj for the next time step will be

�i = JJp(X'Y)�i (x,y)dxdy. (7)

D. Monte Carlo Method for Approximation of Detection Probability

Unfortunately, as shown in Fig. 5, to calculate the joint sensing detection probability analytically for a given emitting sensor is difficult. Therefore, we adopt the Monte Carlo simulation method to generate random samples to approximate the Gaussian distribution in equation (6) of the target location prediction. Fig. 6 illustrates an example for approximating a two-dimensional location Gaussian distribution by random samples. Suppose the total number of samples used is K, then the Gaussian distribution is approximated using discrete probability mass function

(PMF) P(Xj,y) with value 11K for each sample location.

Accordingly, the detection probability Ps of S; is I

approximated as

103

E. Emitting Sensor Selection

In this paper, the emitting sensor is selected as the sensor with the maximal detection probability for individual sensing. After the emitting sensor is selected and activated for emitting wave, for individual sensing scheme only the emitting sensor can take the measurement whereas for the joint sensing scheme, multiple sensors can take the measurements simultaneously.

v. SIMULATION RESULTS

Simulation experiments are conducted for comparing between joint sensing and individual sensing schemes, both based on adaptive sensor scheduling. As shown in Fig. 7, the monitored field is 300cm X 300cm square area. The bottom-left comer of the field is with the coordinate (0, 0), whereas the upper-right comer is with the coordinate (300, 300). Each ultrasonic sensor has the maximal sensing range of 300 cm and the maximal measurement angle is ± 22°. There are eight ultrasonic sensors located along the edge of the area respectively with coordinates (70, 0), (190, 0), (300, 70), (300, 190), (230, 300), (110, 300), (0, 230) and (0, 110). The orientations of the sensors are respectively 90°, 90°, 180°, 180°, 270°, 270°, 0° and 0° such that the field can be covered mostly and any two sensors are not in the detection region of each other. In this setup, each sensor is in the lSI region of any other sensor. We assume the duration of a time slot is 100 ms.

The target moves along a straight line as shown in Fig. 7 with the speed 100cmlsecond. Q is setup at 1.57*1061 where I is the identity matrix. Periodic sensor scheduling is used for initial detection of the target and to initiate the tracking procedure. The initial location estimation of the target is set to the point along the central line of the beam pattern of the detecting sensor with the distance to the detecting sensor equal to the initial measurement. The initial velocity estimation of the target is set to O. The initial covariance can be set heuristically according to the orientation and measurement of the detecting target.

Typical estimated trajectories of adaptive sensor scheduling using individual sensing and joint sensing are shown in the Fig. 7. Clearly, the estimated trajectory by joint sensing is much closer to the true target trajectory as compared with the individual sensing scheme.

The evolutions of tracking errors, i.e., the IQ, are shown in Fig. 8. The gain of the joint sensing is observed. The maximal and averaged tracking error of individual sensing are about 25 cm and 10.01 cm respectively, whereas that of the joint sensing are about 10 cm and 3.66 cm respectively, a significant improvement.

The improvement of the IQ of joint sensing against the individual sensing is due to the increase of the detection region and more simultaneous measurements. Fig. 9 shows the number of sensors used for taking measurements at each time step. Because the individual sensing is a specific joint sensing scenario where the emitting sensor is the same as the receiving sensor, the number of sensors of the individual sensing scheme (being one by default) is always smaller or

Authorized licensed use limited to: Sethu Institute of Technology. Downloaded on July 08,2010 at 05:21:03 UTC from IEEE Xplore. Restrictions apply.

equal to that of the joint sensing scheme. We can fmd that for most of the time steps, the joint sensing scheme has more than two simultaneous measurements except time step 22 and time step 25 with no measurement and one measurement, respectively. However, for individual sensing scheme from time steps 16-24, it does not have any measurements, except time step 23, although one emitting sensor is scheduled for each time step. For time steps 16-21, and time step 22, joint sensing has more than two measurements but individual sensing does not have any measurements, which demonstrates that the joint sensing can increase the detection region significantly.

VI. CONCLUSIONS

A novel collaborative sensing scheme is proposed for target tracking application in WSNs by joint sensing and adaptive sensor scheduling. The proposed scheme can increase the detection region of an individual sensor and introduce more simultaneous sensor measurements for a single sensing operation. It is shown by simulations that the IQ of the WSN can be improved significantly using joint sensing. Future research issues include sensor scheduling for joint sensing for large scale WSNs, adaptive tracking algorithms for high maneuvering targets, joint sensing for multi-target tracking, as well as real test-bed development.

Sensor 6 Sensor 5 3001-------=-::=-::"---=------True target Irajectory

__ Estimated trajectory by individual sensing

250 ---&- Estimated trajectory by joint sensing

Sensor 7 200

150

Sensor 8 100

50

��-·S��n- s�o- r�I , 0� 0-�1�50-Se-n����

�2-�2� 50��300

Figure 7. Estimated trajectories by adaptive sensor scheduling

30

25

'" 20 � Q; � 15 "

� (510

5

� Individual sensing ---e-- Joint sensing

Time steps

Figure 8. Evolutions of tracking errors for adaptive sensor scheduling

104

----ioE- Individual sensing 4.5 ---e- Joint sensing

4

3.5

2.5

1.5

0.5

°0L--��-�1 0�-�1 5�--720�--725�-�30 Time steps

Figure 9. Number of simultaneous measurements obtained by adaptive sensor scheduling

REFERENCES

[I). A. Tolstikov, W. Xiao, 1. Biswas, S. Zhang, and C. K. Tham, "Information Quality Management in Sensor Networks based on the Dynamic Bayesian Network Model," the third International Conference on Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP 2007), Dec 2007, pp. 751-756.

[2). C. Bisdikian, "On Sensor Sampling and Quality of Information: a Starting Point," in Proc. of IEEE PERCOM Workshops, March 2007, pp. 279 - 284.

[3). 1. Wang, Y. Liu, and S. K. Das, "Improving Information Quality of Sensory Data through Asynchronous Sampling," the First International Workshop on Information Quality and Quality of Service for Pervasive Computing (IQ2S 2009) in PerCom 2009, March 2009, pp. 1-6.

[4). F. Zhao, 1. Liu, 1. Liu, L. Guibas, and 1. Reich, "Collaborative Signal and Information Processing: an Information Directed Approach." Proc. IEEE, vol. 91, Aug. 2003, pp. 1199-1209.

[5). W. Xiao, S. Zhang, 1. Lin, and C. K. Tham, "Energy-Efficient Adaptive Sensor Scheduling for Target Tracking in Wireless Sensor Networks," Journal of Control Theory and Applications, Vol. 8, Jan. 2010, pp. 86-92.

[6). W. Xiao, 1. Wu, L., and L. Dong, "Sensor Scheduling for Target Tracking in Networks of Active Sensors," Acta Automatica Sinica, vol. 32, Nov. 2006, pp. 922-928.

[7]. 1. Lin, W. Xiao, F. Lewis, and L. Xie, "Energy Efficient Distributed Adaptive Multi-Sensor Scheduling for Target Tracking in Wireless Sensor Networks," IEEE Transactions on Instrumentation and Measurement, vol. 58, Jun. 2009, pp. 1886-1896.

[8]. L. Chen, B.K. Szymanski, 1.W. Branch, "Quality-Driven Congestion Control for Target Tracking in Wireless Sensor Networks," 5th IEEE International Conference on Mobile Ad Hoc and Sensor Systems (MASS 2008), Sept-Oct 2008, pp. 766 - 771.

[9). W. Xiao, J. K. Wu, L. Shue, Y. Li, and L. Xie, "A Prototype Ultrasonic Sensor Network for Tracking of Moving Targets," the 1st IEEE Conference on Industrial Electronics and Applications (ICIEA 2006), May 2006, pp. 1511-1516.

[10). Y. K. Toh, W. Xiao, and L. Xie, "A Wireless Sensor Network Target Tracking System with Distributed Competition based Sensor Scheduling," the third International Conference on Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP 2007), Dec 2007, pp. 257-262.

[II). Y. Bar-Shalom, X. R. Li, and T. Kirubarajan, Estimation with Applications to Tracking and Navigation. New York: John Wiley & Sons, 2001.

Authorized licensed use limited to: Sethu Institute of Technology. Downloaded on July 08,2010 at 05:21:03 UTC from IEEE Xplore. Restrictions apply.