Research On Multi-target Tracking Based On PHD Filter

Multitarget tracking theory has been widely used in the filed of military and civilian, but it is also the important and difficult problem which is focused by many subjects and fileds. In recent years, the random finite set(RFS) method has attracted more attentions for multitarget tracking and as the approximate product of first-order statistical moment of full multi-target probability density function based on RFS frame, probability hypothesis density(PHD) filter solves the actual execution problem of RFS and it avoids the data association. In this paper, we make deep research on multitarget tracking based on PHD filter. The main contributions are as follows:1. Research on global track extraction based on PHD filter: To solve the problem that it cannot provide the information of continuous tracks of target of PHD filter, global track extraction based on PHD filter is proposed. In the algorithm, it uses the global information of targets, i.e., considers the association of all estimated states at the adjacent two time instants, proposes the concept of consistency measure and consistency belief of predicted peaks and estimated peaks at the same time instant, at the same time, proposes the strategy of global track extraction. In the end, the tracks are extracted one-by-one based on the consistency belief and the constructed decision rules for track extraction and the global track extraction of PHD is implemented. The simulation results indicate that, it can track targets steadily and initialize, maintain, end the tracks correctly of the proposed algorithm. It has obvious advantage on the precision of track extraction and the computation cost is equivalent for comparison.2. The design of improved adaptive PHD filter under the environment of nonhomogeneous clutter and low detection probability: To solve the problem that under the environment of nonhomogeneous clutter and low detection probability, the tracking performance of traditional PHD filter sharply reduces, an improved adaptive PHD filter is proposed. It ensures the fast calculation and the high precision by determining the clutter region adaptively, choosing the measurements adaptively and protecting the Gaussian components with large weights. In the filter, first, cluster for the echoes of several scans satisfying certain condition in the surveillance region using AP clustering algorithm and finding the convex hulls to determine the clutter region. Then, execute PHD predict step and PHD update step for each scan. While executing the PHD predict step, the echoes in the convex hulls are not used. While executing the PHD update step, choose the measurements adaptively first and then execute PHD update. Meanwhile, protect the Gaussian components with large weights and promise their stability of weights. The simulation results indicate that, it can implement target tracking under the environment of nonhomogenous clutter and low detection probability well and comparing with the traditional PHD filter, the estimated precision of target state and the computation efficiency are improved of the proposed filter.3. Research on constrained multiple model PHD filter for maneuvering ground target tracking: Considering that the motion of ground target are restricted by many conditions, such as terrain environment, and so on, it can improve tracking precision effectively to use geographic information for ground target tracking. In the ground target tracking, describe the geographic information as equality constraint to correct the target state, and use the multiple model method to deal with the uncertainty of motion mode while the ground target maneuvers, and the constrained multiple model PHD filter for maneuvering ground target tracking is proposed. In the algorithm, model conditioned distribution and model probability are utilized, each Gaussian component in the frame of the GM-PHD filter is predicted and updated by multiple model method, and the corresponding target states can be gained by fusioning the multiple model estimations. In addition, the road information is described as equality constraint and then it is used to correct the estimated state to complete ground target tracking. The simulation results indicate that, it can estimate the state of the maneuvering ground target with the clutter effectively and comparing it with the MM-GMPHD filter without using the geographic information and the traditional GM-PHD filter, the estimation precision of target state is improved effectively.