Research On The Realization Method Of Probability-likelihood Product For PHD Multi-target Tracking

The essence of information fusion is to synthetically utilize various information on different perspectives and sources to obtain a more accurate and comprehensive understanding and estimation of things.The main research content of this field is to construct and optimize the theory and methods that can achieve the above objectives.Multi-objective tracking is a research hotsp ot and difficulty in the field of information fusion.The probability hypothesis density(PHD)method system based on finite set statistics theory overcomes the predicament that the traditional data association technology brings about that the algorithm pr ecedes the modeling,which makes it possible to construct a unified filtering framework of single target and multi-target.Therefore,it has been proposed by Mahler and other scholars since 2003.It has been receiving sustained and widespread attention.This paper extends the Bayesian filtering probability likelihood product algorithm for single target tracking proposed by the team to the Mixed Gauss scenario and combines it with the PHD system.It provides a new implementation method for the Mixed Gauss Probability Hypothesis Filter(GM-PHD).It solves the problem that GM-PHD filter cannot deal with the non-linear multi-target tracking system and multi-maneuvering target tracking under dense clutter.The main contents are as follows:1.The theoretical basis of random finite sets is expounded,in which the optimal multi-objective Bayesian filter and its approximate filter method probability hypothesis density filter are introduced,and approximate realization form of probability hypothesis density filter is given.The advantages and disadvantages of this realization form are compared with Monte Carlo realization method.2.In order to overcome the difficulties of realiz ing and the low accuracy of target state and number estimation of GM-PHD filter in non-linear system and clutter environment,a new kind of PDH filter based on probability likelihood product filter is proposed.On the basis of Bayesian theory,a new probability likelihood product fusion filtering rule is deduced.Based on this rule,a more w idely used probability likelihood product filtering algorithm is obtained.In the non-linear non-Gauss system,the tracking performance of this algorithm is excellent for single target.Combined with PHD filter,the probability likelihood product filter is extended to multi-target tracking,and the product fusion filter is updated for each Gauss component.The simulation results show that,compared with the existing non-linear GM-PHD filtering algorithm,this algorithm is more convenient in engineering application and has better tracking performance.3.Aiming at the problem of low tracking performance and inaccurate estimation of target number caused by target maneuvering in multi-target tracking,based on the idea of interactive multi-model,an IMM-PHD algorithm of probability likelihood product is proposed.In the framework of GM-PHD filtering,IMM algorithm is used to recognize the motion model,and probability product rule is used to filter the Gauss component of single model.Finally,multi-model target state is fused and estimated.The simulation results show that the proposed algorithm has high stability and performance.