Research On Multi-target Tracking Algorithm Based On Probability Hypothesis Density Filter

As a significant part of information fusion technology,multi-target tracking technology has been widely used in various civil and military applications.With the increasing complexity of tracking scenarios,conventional multi-target tracking algorithms based on data association are burdened with high computational complexity because of complex data association.In recent years,the probability hypothesis density(PHD)filter has been developed to avoid the complex data association.In this way,the PHD filter has low computational complexity and is easy to implement.This thesis mainly focuses on the research of PHD filter based multi-target tracking algorithms.The purpose of this thesis is to solve the following problems: point target tracking in complex scenarios,measurement set partition in extended target tracking,and the shape estimation of extended target.The major contributions of this thesis are summarized as follows:1.A robust sequential Monte Carlo PHD filtering algorithm for multi-target tracking is proposed with measurement noise outliers.Firstly,the Student-t distribution with unknown covariance matrix and degree of freedom(DOF)is introduced to describe the measurement noise.And then,the Inverse Wishart distribution and the Gamma distribution are utilized to model the covariance matrix and DOF of the Student-t distribution,respectively.Furthermore,the Variational Bayesian method is employed to infer the unknown noise parameters while the implementation of the robust SMC-PHD filtering algorithm is given.In addition,the particle weight is modified in the update step,which solves the overestimation of the target number caused by the Student-t distribution.Simulation results show that the proposed algorithm can deal with the multi-target tracking problem with measurement noise outliers.2.Considering the problem of wrong estimation,false estimation,and missed estimation when the collaborative penalized Gaussian Mixture PHD(CPGM-PHD)filtering algorithm is used to track multiple spatially close targets,an improved CPGM-PHD filtering algorithm is proposed.Firstly,aiming at the wrong estimation for CPGM-PHD filtering algorithm when targets are spatially close,a weight rearrangement scheme is proposed to rearrange the weights of Gaussian components assigned to each target.Then,using the continuous property of the target trajectory,the missed target at the current time is modified by the predicted value at the last time to reduce the missed estimation.Finally,the estimated targets are classified by making full use of the multi-frame estimated target states,and the false estimation is detected and deleted.Simulation results show that the improved algorithm has higher estimation accuracy.3.The measurement set partition in extended target tracking is studied,and measurement set partition algorithm based on the clustering by fast search and find of density peaks(CFSFDP)is proposed.In the proposed algorithm,the CFSFDP method is introduced into the measurement set partitioning.Considering that the cutoff distance parameter and cluster center of CFSFDP method need to be determined manually,the data field method is further introduced to determine the cutoff distance parameter adaptively,and then the cluster center is also adaptively determined according to the solved cutoff distance.Simulation results show that extended target Gaussian Mixture PHD(ET-GM-PHD)filtering algorithm,which uses the proposed algorithm to partition the measurement set,has lower computational complexity.In addition,to solve the problem that the existing partitioning algorithm is difficult to partition the measurement set when targets with different sizes are spatically close,an adaptive measurement set partitioning algorithm is proposed.In the proposed algorithm,we first perform the preliminary partition of the measurement set using the aforementioned CFSFDP-based partition algorithm.In view of the sub-partitioning problem of the measurement set,we improve the determination conditions of sub-partitioning and traditional Kmeans algorithm.The Mahalanobis distance is used to replace the Euclidean distance so that the effect of the target size can be reduced.Simulation results show that the extended target Gaussian Inverse Wishart PHD(ET-GIW-PHD)filtering algorithm that uses the proposed algorithm to partition the measurement set has higher estimation accuracy when targets are spatially close.4.When the attitude of the target changes,the traditional method that uses a fixed number of sub-ellipses to describe the shape of the target will lead to a large estimation error in target shape.To solve this problem,an extended target tracking algorithm is proposed with the varying number of sub-ellipses.Firstly,the target shape is described by multiple spatially close sub-ellipses and the target labels are introduced to label the target.The sub-ellipses targets of the same extended target have the same labels,and sub-ellipses targets of different extended targets have the different labels.The association is realized between the extended target and the corresponding sub-ellipses targets.Then,by target spawning and combination,the number of sub-ellipses for approximating the shape of target can be adjusted automatically.Finally,the aforementioned adaptive measurement set partition algorithm is improved to partition the measurement set.Simulation results show that the proposed algorithm can accurately estimate the number of sub-ellipses approximating the shape of the target and has higher estimation accuracy when the shape of the target changes.