Multi-target Tracking Algorithm Researches Based On Probability Hypothesis Density

In recent year, Probability Hypothesis Density (PHD) algorithms arereceiving great interest in multi-target tracking since it can easy handlemulti-target tracking with unknown number of target, uncertainty associatedbetween target and measurement problems. In addition, highly nonlinear andnon-Gaussian estimation problems are ubiquitous in multi-target tracking, andParticle implementation of PHD (P-PHD) is an effective tool for such problems,therefore, in this dissertation, a key technique in P-PHD, multi-individual targetsand multi-group targets are studied in the framework of P-PHD. The maincontributions are as follows:P-PHD is a multi-target tracking algorithm without data association operation,it is good at handling multi-individual and multi-group target tracking in highclutter. In the filtering processing, the information of target number and state areincluded in the intensity function characterized by a group of weighted particles.When extracting the target state, we need to utilize EM or k-means algorithm tocluster the resampled particles into different parts and estimate the state of eachtarget, but, the computation burden on those clustering algorithm are very large. Inaddition, P-PHD algorithm will suffer such as the reduction of the diversity sinceits particle-implementation method, which will affect the estimation accuracy ofstate and number of target, moreover, the higher the dimension of target state, themore impact on the performance of P-PHD algorithm. Therefore, the group targetalgorithm is more sensitive to the lack of diversity of particle, because it needs toestimate not only the centroid of group target, but also the shape of group target.To lessen the computation burden on extract target state and enhance thediversity of particle, according to the characteristics of multi-individual target andmulti-group target tracking, this paper presents two corresponding improvedP-PHD filter.1) Firstly, according to the analysis of the principle of resamplingprocessing in P-PHD, we find out the similarity between the clustering algorithmand basic resampling processing, and present employing the basic resampling algorithm to cluster the updated particles into different parts based on theobservation information. Then, utilizing the UKF to move each classified particleinto high likelihood area based on its associated measurement, by doing so, thediversity of particle is increased, and the accuracy estimation of state and numberof target is enhanced.2) Aiming to multi-group target, we firstly view eachreflection dot, which generates effective observation, as the individual target.Then, the algorithm adopts the proposed methods, aimed to multi-individual target,to extract the state of each reflection dot. At last, the shape of group target isestimated based on the information of the reflection dot.