A Practical Target Tracking Technique in Sensor Network Using Clustering Algorithm

Sensor network basically has many intrinsic limitations such as energy consumption, sensor coverage and connectivity, and sensor processing capability. Tracking a moving target in clusters of sensor network online with less complexity algorithm and computational burden is our ultimate goal. Particle filtering (PF) technique, augmenting handoff and K-means classification of measurement data, is proposed to tackle the tracking mission in a sensor network. The hand-off decision, an alternative to multi-hop transmission, is implemented for switching between clusters of sensor nodes through received signal strength indication (RSSI) measurements. The measurements being used in particle filter processing are RSSI and time of arrival (TOA). While non-line-of-sight (NLOS) is the dominant bias in tracking estima-tion/accuracy, it can be easily resolved simply by incorporating K-means classification method in PF processing without any priori identification of LOS/NLOS. Simulation using clusters of sensor nodes in a sensor network is conducted. The dependency of tracking performance with computational cost versus number of particles used in PF processing is also investigated.


Introduction
Sensor networks can be applied in a variety of areas such as target tracking, environment monitoring, military surveillance, medical applications, etc. [1,2].However, the majority of sensor node platforms are operated using the low power 802.15.4 wireless technology, and its transmission range is extremely limited especially in an indoor environment [3].The measurements used in the estimate of mobile locations in sensor network may include received signal strength, time difference of arrival (TDOA), time of arrival (TOA), and angle of arrival (AOA) [4].Eventually, the above propagation measurement scenarios are divided into two categories, line-ofsight (LOS) and non-line-of-sight (NLOS).In multi-path propagation environments, particularly indoors or urban areas, the LOS path between nodes may be obstructed [5].However, the NLOS propagation usually leads to a positive bias and causes a serious error in the results of tracking estimation [6].
Lots of attentions have been focused on the identification of LOS/NLOS condition and the mitigation of NLOS bias.A simple hypothesis test has been conducted to tell whether it's LOS or NLOS by the fact that the standard deviation of the range measurement of NLOS is presumably larger than the LOS' [6].Using the individual measurement detection (IMD), basically a hypothesis test to identify whether an incoming measurement is LOS or NLOS and those NLOS ones being discarded, to do target tracking is proposed [7]; extended Kalman filter (EKF) algorithm is applied accordingly to do the target tracking job.The noise modeling of LOS/NLOS is formulated by a two-state Markov process, and the degree of contamination by NLOS errors is correlated with the transition probability of the Markov process.A disadvantageous effect has been indicated, the number of selected LOS measurements by IMD is different at each step, resulting in dimension validation of the reconstructed LOS measurement vector being dynamic; slow convergence rate is also appearing in the tracking results when using EKF.A modified Kalman filtering technique, adopting the modification at the measurement update stage, is introduced to tackle the NLOS identification/mitigation problem [8].The NLOS positive bias is estimated directly throughout the constrained optimization method; no prior distribution knowledge of the NLOS error is needed, as claimed by the authors.
Our proposed tracking algorithm utilizes clusters of sensor network with handoff scheme in a heterogeneous wireless environment.A handoff scheme specified in IEEE 802.11 network is the process whereby a mobile station shifts its association from one access point (AP) to another [9].When a mobile station moves out of the range of an AP and into another's, the handoff occurs during which there is an exchange of management frames [10].
On the other hand, clustering analysis is the method of unsupervised learning.It may include two parts, namely, partitional clustering and hierarchical clustering.Partitional clustering is a set of objects classified into K clusters without hierarchical structure [11].The clustering method with PF can reduce the degradation of NLOS propagation efficacy in tracking estimation results.Meanwhile, handoff scheme is applied in the clusters of sensor network.In clusters of sensor network, each cluster has numbers of sensor nodes, including TOA and RSSI sensors.Figure 1 shows the proposed architecture that illustrates an event of target tracking in clusters of sensor network.
The general approach to processing the LOS/NLOS signals in cellular communication network is to determine whether it's a LOS or NLOS condition.Even the fuzzy inference scheme is introduced to tell whether it is LOS or NLOS before any processing jobs can be done [10].It is combined with adaptive Kalman filter to establish mobile location estimator; undoubtedly, system and computational complexities are increasing so tremendously, hindering the potentials in real-time applications.
We present an architecture which utilizes clustering analysis method with particle filter (PF) to track a moving target.Particle filter implements sequential Monte Carlo simulation based on a set of particles to construct prior density with associated weights for the approximation of posterior density.
The beneficial features of the proposed scenario are listed as bellows,  Intuitive but feasible for real-time applications due to its low system and computational complexity attributes  No need to identify whether it is under LOS or NLOS condition prior to any processing of the target location estimation  Only RSSI and TOA sensor measurements are required, practical and low-cost; fingerprinting scheme, the build-up of a "radio map" database [6], and extra hardware (sophisticated measurement devices) are not employed, always leading to a low complexity, costeffective scenario. Both K-means clustering and hand-off schemes are incorporated into the particle filter, minimizing the system complexity to the most.This analysis is set up as follows.The target's motion model and the associated measurement equation and NLOS propagation are described in Section 2. Particle filter is described in Section 3. Section 4 discusses the proposed tracking algorithm which includes the handoff scheme.The proposed tracking algorithm with clustering method is derived in Section 5.The simulation and performance analysis are presented in Section 6. Section 7 includes the conclusion.

Motion Models
We assume a mobile target's movement is on a two-dimensional (2-D) plane.Besides, the measurement of signal propagation with LOS/NLOS conditions may be modeled as a two-state Markov process if the performance at transient stage is under investigation.

Target Model
The moving target's state vector is defined as , consisting of position and velocity at a time instant k, where () T stands for transpose operation of matrix.The target's motion is modeled as and I 2 is the 2  2 identity matrix;  is the Kronecker product operator; A 1 is the state transition matrix; s T is the sampling time; and is a zero mean white Gaussian noise process with covariance matrix Q 1 , i.e., where q 1 is a scalar which determines the intensity of the   As the simulation results shown from Figure 12 to Figure 14, each figure is simulated with 1000 time instants, and we compute its associated estimation errors, position, velocity, and acceleration.
As position, velocity, and acceleration are estimated, not too much benefit can we expect in a LOS propagation environment when we vary the number of particles in PF processing.The number of particles we choose is 50, being attractive in real-time tracking applications.The estimation errors gradually reduce with the increaseing of particle numbers.On the contrary, situated in a NLOS environment, substantial improvement of estimation errors are illustrated in Figure 14 with K-means clustering scenario; eventually, the trade-off among the increment of particle numbers, estimation errors, and e particle filter processg job.

Conclusions
Sophisticated and high system/computational complexity algorithms are always proposed to mitigate the NLOS effect and estimate the mobile/target location.In this article, we propose a simple and feasible generic tracking algorithm to track the moving target in clusters of sensor network.The proposed tracking algorithm is the technique that adds handoff decision to the ordinary tracking algorithm, based on TOA and RSSI measurements; the handoff decision is implemented on clusters of sensor network.
Besides, K-means clustering is utilized, and it combines with particle filter to reduce the NLOS propagation effect.Finally, the proposed algorithm can accom [1] P. M. Djuric, M. Vemula and M. F. Bugallo, "Tracking computational load is accomplished via the use of moderate number of particle, i.e., 50, and the augmentation of K-means clustering scheme in th in plish higher accuracy in tracking estimation for sure.
Simulations illustrate that the estimation results of tracking trajectory is well predicted, even around the NLOS propagation environment.This analysis applies to any motion modes, even with varying acceleration.Moreover, we also compare the results of tracking algorithm with and without K-means clustering in statistics.Through the performance analysis, it demonstrates that the proposed tracking algorithm may find potentials in real-time tracking/localization applications as the particle numbers used are reducing to as low as 50.

Figure 1 .
Figure 1.The proposed architecture in clusters of sensor network.

Figure 12 .
Figure 12.The results, estimation errors versus numbers of particles, using solely particle filtering in a surrounded LOS propagation environment.

Figure 13 .
Figure 13.The results, estimation errors versus numbers of particles, using solely particle filtering in a surrounded NLOS propagation environment.

Figure 14 .
Figure 14.The results, estimation errors versus numbers of particles, using particle filtering with K-means clustering in a surrounded NLOS propagation environment.