Poisson Multi-bernoulli Mixture Filter And Its Implementation For Multiple Extended Target Tracking

The target is treated as a point in traditional multi-target tracking.With the improvement of the resolution of sensors such as radars,multiple measurements will be obtained by a detected target,so it is necessary to treat the point target as an extended target.While tracking its centroid,the specific information such as the extended shape and diffusion range of the target is also estimated,so as to be able to track and recognize the target more comprehensively.At present,extended target tracking has been increasingly used in aerospace,indoor and outdoor positioning,artificial intelligence and other aspects.One main method for multiple extended targets tracking is to apply a tracking algorithm based on Random Finite Set(RFS).In recent years,multiple extended targets tracking based on RFS has been rapidly developed,and more and more filter algorithms related to have been proposed.The Poisson Multi-Bernoulli Mixture(PMBM)filter is one of them,which models the state of multi-target as two independent parts of Poisson and Multi-Bernoulli Mixture,which is used to to represent the detected targets and the undetected targets.The two parts are predicted and updated respectively.The algorithm has the advantages of high tracking accuracy and fast calculation speed.This thesis mainly focuses on the depth study of the PMBM in multiple extended targets tracking and the specific research contents are as follows:(1)Based on the theory of extended target tracking,combined with particle filter and box particle filter respectively,the extended target sequential Monte Carlo PMBM(ET-SMCPMBM)filtering algorithm and the extended target box particle PMBM(ET-BP-PMBM)filtering algorithm are proposed,and the algorithm flow of the two algorithms are given.The simulation results show that the extended target PMBM filter has higher tracking accuracy than the Probability Hypothesis Density(PHD)filter.Compared with the ?-Generalized Labeled Multi-Bernoulli(?-GLMB)filter,the two filters have similar tracking results but the calculation speed of the PMBM is faster than the ?-GLMB.ET-BP-PMBM will cause a certain error accumulation due to the box particle filtering algorithm using interval analysis method,so the tracking accuracy of the proposed algorithm is decreased,but the operating efficiency is higher and the tracking speed is faster.(2)For the shape estimation of the extended target,the sequential Monte Carlo PMBM filtering algorithm based on the PMBM filter and random hypersurface model is proposed,and which has realized the estimation of the shape with ellipse.Finally simulation experiments show that the algorithm can adapt to complex tracking scenarios and has high accuracy of shape estimation for extended target.(3)The three performance evaluation metrics in multi-target tracking based on random finite sets : Optimal Subpattern Assignment(OSPA)distance,Generalized Optimal Subpattern Assignment(GOSPA)distance and Complete Optimal Subpattern Assignment(COSPA)distance have been introduced.The meanings,characteristics and usage scenarios of three metrics are respectively summarized and three metrics are applied and analyzed in the simulation experiments.