Mcmc Particle Filter Using New Data Association Technique with Viterbi Filtered Gate Method for Multi-target Tracking in Heavy Clutter

— Improving data association technique in dense clutter environment for multi-target tracking used in Markov chain Monte Carlo based particle filter (MCMC-PF) are discussed in this paper. A new method named Viterbi filtered gate Markov chain Monte Carlo VFG-MCMC is introduced to avoid track swap and to overcome the issue of losing track to highly maneuvering targets in the presence of more background clutter and false signals. An adaptive search based on Viterbi algorithm is then used to detect the valid filtered data point in each target gate. The detected valid point for each target is applied to the estimation algorithm of MCMC-PF during calculating the sampling weights. This proposed method makes the MCMC interacts only with the valid target that is candidate from the filtered gate and no more calculations are considered for invalid targets. Simulation results demonstrate the effectiveness and better performance when compared to conventional algorithm MCMC-PF. I. INTRODUCTION In multi-target tracking systems (MTT) data association is an important key to associate the track to the true target in noisy measurements. The measurements (target of interest, clutter, noise signal) are reported by sensor at regular time interval (scan period). The received measurements in a cluttered environment may not originate from the real targets. Some of them may be from clutter or false alarm. As a result, the association between the tracked target and true candidate measurement is difficult. Assigning wrong measurements to track often results in lost track and track breaks. Moreover, in heavy clutter density the resulting number of false tracks may overwhelm the available computational resources of the MTT systems. For this reasons, improving data association technique has received much attention in MTT research [1-6]. The target tracking consists of two basic parts: data association and tracking filtering. Data association is responsible for detecting on each scan which of the received multiple measurements that


I. INTRODUCTION
In multi-target tracking systems (MTT) data association is an important key to associate the track to the true target in noisy measurements.The measurements (target of interest, clutter, noise signal) are reported by sensor at regular time interval (scan period).The received measurements in a cluttered environment may not originate from the real targets.Some of them may be from clutter or false alarm.As a result, the association between the tracked target and true candidate measurement is difficult.Assigning wrong measurements to track often results in lost track and track breaks.Moreover, in heavy clutter density the resulting number of false tracks may overwhelm the available computational resources of the MTT systems.For this reasons, improving data association technique has received much attention in MTT research [1][2][3][4][5][6].The target tracking consists of two basic parts: data association and tracking filtering.Data association is responsible for detecting on each scan which of the received multiple measurements that lie in the specified gates of the predicted targets position should update with the existing tracking targets.Due to data association, each scan takes the output of the gating process [7] and makes final measurement to track associations.A gating process with filtering method based on Viterbi algorithm is used in this part to find the valid measurement that is considered as the best candidate measurement originated from the true target and assume the remaining measurements to be invalid that originated from false alarm or clutter.The Viterbi algorithm (VA) [8] is a dynamic programming technique for finding the shortest path through a trellis in a computationally efficient manner.The using of the Viterbi algorithm for solving the data association problem when tracking targets in clutter was first suggested as VA-QF by Quach and Farooq [9].Their approach finds the approximate maximum likelihood assignment of measurements to the target.The problem they considered was that of tracking a single, maneuvering target assumed to be present in surveillance region.This method was modified in [10] to suit the problem of tracking multiple, non maneuvering targets in heavy clutter and was applied to tracking with an Over-The-Horizon radar (OTHR).Instead of using [9] in data association for single target that deals with multi state (multi-scan) where each state contains all measurements according to the corresponding time scan and is processed with the next scan up to all measurements received in the current scan.A new technique for gating process using Viterbi algorithm for multi-target tracking is introduced to be considered as an extension of [9] to detect the measurement in the current gate of the tracked target that has the maximum cost.This extension method for each track depends only on the measurements in current and previous gate of the same tracked target position by two successive states, each state represent the corresponding time scan for all tracked target gate.The candidate measurements in data association part are applied to tracking filter part for estimation.Due to the track swapping and highly maneuvering target, the mainly considered problem for estimation, traditionally, is solved using linearized filters, such as the extended Kalman filter (EKF) [11][12][13][14][15], under a www.ijacsa.thesai.orgGaussian noise assumption.The sufficient statistics from the linearized filter are used for data association.When dealing with non-linear models in state equation and measurement relation and a non-Gaussian noise assumption, these estimation methods may lead to non-optimal solutions.The sequential Monte Carlo methods, or particle filters [14][15][16][17][18][19][20][21], provide general solutions to many problems where linearizations and Gaussian approximations are intractable or would yield too low performance.Recently MCMC-based particle filter (MCMC-PF) [22][23][24][25][26] has captured the attention of many researchers in various applications that deal with difficult nonlinear and/or non Gaussian problems for tracking maneuvering target in clutter.In MCMC-PF, the particles are sampled from the target posterior distribution via direct MCMC sampling method, which avoids sample impoverishment.To increase the robustness of the algorithm and tracking a highly maneuvering target without swap in dense clutter environment, we introduce the proposed algorithm named VFG-MCMC.The new technique based on Viterbi algorithm as mentioned above is combined with MCMC-PF.Simulation results showed better performance and more effective at tracking when compared to the conventional MCMC-PF algorithm.

II. PROBLEM FORMULATION
In a dynamic state space model, the observed signals (observation/ measurements) are associated with a state and measurement noise.Consider that there are T targets being tracked at time index k.The time evolution of the state of a discrete-time, nonlinear, dynamic systems is described by the following equation where f is the system transition function and R n k w t   ) 1 ( is a dynamic noise which has a known probability density function (PDF).The superscript t corresponds to the th t target.
The initial target state,   0 x t for t = 1,2, ..., T , is assumed to be known.At discrete times, the measurements become available and are related to state through the observation equation . . . . . . . . . . . .

 
where h is a nonlinear measurement function and ) ( is a sequence of observation noises of known PDF.From the Bayesian perspective, the tracking problem is to recursively calculate the posterior probability density function,

A. Viterbi Data Association VA-QF for Single Target
The basis of the VA algorithm is built upon the Markov process [9], [27], [28].In VA algorithm, the target motion is assumed to be a Markov process.The Markov process uses state diagrams where the nodes or measurements represent the state in each scan and the arcs are represented by transition paths.As time progress, the process will trace some path from state (or scan) to state through the state diagram.If the states and their transitions are repeated in time for k=1,..,K, we obtain a trellis that represent a directed graph consisting of a set of measurements and paths as shown in Fig. 1.At each measurement from scan to scan, the length of the paths coming into the measurement is summed.The path with the shortest length will be assigned to that measurement and is considered as survivor path (one survivor for each measurement).the recursion will proceed from first to last (current) scan, once the last measurement is reached, the algorithm summing all the smallest (path metric) nodes from each previous scan to obtain the optimal path.A path is a collection of directed arcs that connects an element at stage 1 to an element at stage K.Each directed arcs is associated with a metric or a distance label . A path metric is defined as the sum of the metrics of all the arcs contained in the path as .
By applying the Viterbi algorithm to the trellis diagram, we obtain the sequence of measurements that maximizes the overall likelihood function, thus solving the data association problem as in the following steps.1) Initialization step: assign a value of zero to the label at each node/measurement in stage 1: where the innovation with the smallest label as follow: and let  update the target state ) ( ˆk x j and its covariance at each node in stage k 3) Final selection: determine the node with the minimum score in the final stage.The current state of the target is the state associated with the minimal node.
4) Back track: recover the measurement sequence that terminates with the minimum node score in the final stage using

B. Markov Chain Monte Carlo based Particle Filter(MCMC -PF) 1) Basic theory of particle filter
The particle filter is a recent filtering method based on Bayesian estimation and Monte Carlo method and effectively deals with complicated nonlinear and /or non-Gaussian problems.The basic idea of this method is Monte Carlo simulation, in which the posterior density is approximated by a set of particles with associated weights At every time step we sample from the proposal distribution to achieve new particles and compute new weights according to the particle likelihoods.After normalization of weights, the posterior density can be represented by   The implementation of the basic particle filter is as follows: (1) Initialization Draw a set of particles from the prior : Prediction: (2) Sampling step Update: (b) Evaluate the importance weights when the candidate measurement ) (k z is choosing by data association algorithm (c) Normalize the weights (3) Markov Chain Monte Carlo (MCMC) step as described in algorithm 1 can be used to approximate the posterior distribution. Expectation: Covariance: (5) k=k+1, go to step 2

2) Basic theory of Markov Chain Monte Carlo (MCMC)
The MCMC step, as described in [24] has an invariant , which is applied to each of the N particles, one at the time.

III. VITERBI FILTERED GATE DATA ASSOCIATION
To introduce a robust filtering technique with improving tracking performance and give the tracking system the capability to track the multi-target when moving with highly maneuvering without swapping and in dense clutter environment, a new algorithm, named Viterbi filtered gate based Markov Chain Monte Carlo-particle filter (VFG-MCMC) as shown in Fig. 2, is proposed.
whose elements are a subset from the set where i =1 to mi ( number of detected points in gate for each target t where  is threshold value that determines the gate size and l =1 to w n , i =1 to mi, i. e for each target t, i is initialized by 1 and is increased by i = i +1 after each valid point is detected up to last mi detected points.
In the updating step, let for each target t where l =1 to w c , j=1 to mj for j=j+1 after each valid point is detected.A new filtering gate method based Viterbi algorithm is proposed to distinguish between the detected measurements in that originated from the target or originated from clutter ( false target ).
In this method we consider the two constructed gate for each target as the two successive state as mentioned above in viterbi algorithm VA-QF for single target.This method is based only on the measurements (nodes) that fall in the two previous and current gate ,   k G t } for each target that have the same predicted target position at time scan k-1, k respectively.www.ijacsa.thesai.org to be considered as the valid measurement that originate from the target and as the choosing measurement in the updating step, we first calculate distance metric between each measurement where the innovation in this method is calculated as The minimum distance d t j min that represents the shortest path at each measurement ) (k z j is assigned to its to give For each measurement label j in that has the minimum distance and then its measurement label (index) After determining the shortest path with minimum distance , the valid measurement * j that is a candidate to be the associating measurement to a tracked target t in the updating step is calculated by obtaining the measurement index label of the maximum ) (k score j as in the following equation Thus, for each target t at each new scan, the filtered gate using this proposed method of data association technique select one measurement from the target gate that is considered to be valid (originate from the true target) and is assigned to be the selected or candidate measurement used in the process for estimation in the updating step, while the remaining measurements are considered to be invalid (originate from clutter) and are avoided from tracking process.

IV. IMPLIMENTATION OF VITERBI FILTERED GATE DATA ASSOCIATION WITH MCMC-PARTICLE FILTER (VFG-MCMC)
The proposed algorithm VFG-MCMC is an approximate likelihood-based approach for solving the data association problem in multi target tracking when the clutter is heavy.This new algorithm depends on the history of measurements for one scan and the construction current gate.The current gate of the tracked target is processed with the previous gate of the same tracked target position using an advanced technique based on the basic principle of viterbi algorithm VA-QF.This technique detects the one valid measurement in the current gate that is assumed to be originated from true target, while the remaining measurements in the gate are considered as invalid that are assumed to be originated from clutter or false target.This filtering gate method select or candidate the one valid measurement for each target to be processed in the updating step while the invalid measurements not include in the tracking process.The proposed data association algorithm is combined with MCMC-PF to overcome the problem of highly maneuvering target and swapping in the presence of dense clutter.The proposed algorithm VFG using MCMC-PF is represented in algorithm 2.

For t=1 to
By accepting only those measurements that lie inside the gate t: 2. Find validated region for measurements at time k: By accepting only those measurements that lie inside the gate t 3. For each measurement j the distance metric a k ij is calculated by ( 22) using ( 23), (24) 4. The distance path ) (k a ij  is obtained by (25) 5.For each measurement j finds the label measurement * i as calculated in ( 26) that has the minimum distance a j i * 6. Find the valid measurement * j that has the maximum score Using (27).
Finally, by using VFG data association algorithm ,we obtain for each target the valid measurement in the tracked target gate that is applied to MCMC-PF before entering in the process of the updating step .The other measurements in the gate are assumed to be invalid and the updating process for tracking the targets is not assigned to any one of them.VFG data association algorithm has an advantage that is accepted by MCMC-PF to improve the performance in heavy clutter for the tracked targets and to maintain the tracking to the targets that move with high maneuvering and swapping.

V. SIMULATION RESULTS
In order to demonstrate the performance of the proposed VFG-MCMC algorithm, we provide a comparison with the most popular nonlinear conventional tracking algorithm MCMC-PF.Two test scenarios have been chosen in this section for moving more than one target in the XY plane with different issues according to the two different tracking problems; high maneuvering and swapping in dense clutter environment.The discrete time state equation with sampling is the state vector ; ) (k x and ) (k x  are respectively the position and velocity of the moving object along the Cartesian frame X axis; and . The measurements are received from one sensor which is positioned at the origin of the plane.The measurement equation is as follow: , where is the distance between the origin and the moving target and ) ( 2 k z is the bearing angle.
The measurement noise is a zero mean Gaussian white noise process with variance R , where is the position of the sensor at the origin, ), we showed that the proposed VFG-MCMC succeeded to track the targets in the same environment as shown in Fig. 3(c).We obtain trajectories for X-and Y-components as shown in Fig. 4(b),(c).In these figures, the colored solid line represents the underlying truth targets of the trajectory (each target with different color) while the colored + symbol represents trajectory of the estimated tracked targets.Our proposed algorithm (+ symbol with different color) detects and associates the proper sequence of observation very well compared to MCMC-PF which fails to continue.According to the second test scenario, the conventional algorithm in clear environment with no clutter fails to avoid swapping between tracks as shown in Fig. 5(a),(b),(c),(d) that are represented in different time scan and also its X-Y-trajectory is implemented shown in Fig. 6 .Thus, when using the conventional algorithm for tracking the targets in more background clutter we fail to track the targets with no avoiding the swapping of the crossed targets as shown in Fig. 7(a).The proposed algorithm that is evaluated in the same clutter environment succeeded to track the targets with no swapping as shown in Fig. 7 (b).The X-Y-trajectories for the two approaches to demonstrate the swapping effect and the capability to maintain the tracks in more background clutter are shown in Fig. 8

VI. CONCLUSION
In this paper for multi-target tracking in dense clutter environment, a new data association technique based on filtering gate method using Viterbi algorithm (VFG) applied to MCMC-PF, can solve the tracking problem effectively, such as ; highly maneuvering target, track swapping.From the results obtained in the simulations, we have showed that in dense clutter environment, the MCMC-PF fails to tracks the high maneuvering targets and fails to avoid track swapping for the moving crossed target trajectory , where the proposed Viterbi filtered gate (VFG-MCMC) algorithm is capable of tracking the targets and avoid swapping.The VFG-MCMC avoids the false targets from the valid based measurement regions and chooses the best candidate valid measurement to the tracking filter MCMC-PF using VFG algorithm.Thus, this improves the data association process which has been shown to give targets the ability to continue to be tracked in dense clutter and also improves the tracking performance of MCMC-PF.This approach can be used to overcome the clutter of gate based approaches in tracking.

Figure 2 .
Figure 2. Schematic Diagram of VFG data association algorithm To detect the candidate valid measurement ) ( * k z j from the For each node j detect the node i at that has the minimum distance with path of node j , 200m/s, 200m/s], which continues from the first frame to the last frame.This scenario is used to evaluate the tracking performance by the proposed and conventional algorithm when the targets move with high maneuvering in the presence of heavy clutter.The second test scenario has two tracks with initial state ) 0 ( 1 x = [13.4km,12.5km, 150m/s, 150m/s], ) 0 ( 2 x = [13.3km,10.5km, 150m/s, 150m/s]moving in cross trajectory corrupted by more background of clutter to evaluate www.ijacsa.thesai.orgthetracking performance against swapping between tracks for the two approaches that are presented here.We initiate the other parameters as; the row and column sizes of the volume ( first test scenario, using the conventional MCMC-PF algorithm in clear environment with no clutter to track highly maneuvering targets, continuing to tracking with no failure as expected as shown in Fig.3(a).The X-Y trajectory is implemented in Fig 4(a).But this algorithm at low signal to noise ratio fails to track the corrupted targets with a uniform high clutter density as shown in Fig. 3(b) while, given a fixed threshold (

Figure 3 .Figure 4 .Figure 5 .Figure 6 .Figure 7 .
Figure 3.The state of tracking multi-targets (3 moving targets) using the 2 approaches (a) MCMC-PF succeeded to track in clear environment (no clutter) (b) MCMC-PF failed to continue tracking in heavy clutter (c) VFG-…MCMC succeeded to continue tracking in the same environment as in (b) Figure 8. X-Y Components Trajectory for the 2 approaches algorithm used in tracking 2 targets (+ symbol) in more background clutter and showed the true target path as in solid line.(a) Trajectory for X-Y by MCMC-PF. . . .(b) Trajectory for X-Y by VFG-MCMC.

Figure 9 .
Figure 9.The RMSE for each target (3 targets) separately over frame number (each frame take 4 sec / one scan) according to tracking evaluation of the first .test scenarios as shown in Fig. 3,4 .

Figure 10 .
Figure 10.The RMSE for each target (2 targets) separately over frame number (each frame take 4 sec / one scan) according to tracking evaluation of .the second test scenarios as shown in Fig. 5,6,7,8.