Probabilistic Graphical Model Based Multiple Target Tracking Algorithms

Multi-target tracking is one of the relevant research topics in the field of information fusion.It has been widely used in military fields such as air surveillance and early warning,missile defense,battlefield surveillance,and civil fields such as computer vision,air navigation,traffic control.However,with the rapid development of high-resolution sensors and the increasingly complex battlefield environment,the obtained measurements of targets are more abundant.The data association problem in multi-target tracking becomes more complicated.Meanwhile,the latest random finite set(RFS)filters have involved the data association problem.Solving the data association problem with high efficiency is needed.Therefore,efficient and robust multi-target tracking technology is a hot issue in the field of target tracking.Recently,considerable attention has been attached to multi-target tracking based on the probabilistic graphical model theory.Probabilistic graphical model is an effective inference and learning tool in large scale systems.As a new approach to solve multi-target tracking in complex environment,the data association problem in multi-target tracking can be modeled by a probabilistic graphical model.In this way,the belief propagation algorithm can efficiently solve the data association problem in multi-target tracking.Meanwhile,the probabilistic graphical model can also be used to improve the computational efficiency of RFS filters.Based on the probabilistic graphical model theory,the research of this dissertation puts its emphasis on multiple extended target data association methods and RFS filters with high efficiency and robustness.The main contributions of the dissertation are given as follows:1.The problem that the marginal distribution of association variables in multiple extended target tracking is difficult to obtain has been studied.Because one extended target generates more measurements than one point target at each time step,the data association problem in multiple extended target tracking is more complicated.The computational complexity of traditional data association methods is too high to implement.In order to solve this problem,a data association algorithm for multiple extended target tracking is proposed based on a track-oriented probabilistic graphical model.In the proposed algorithm,the data association problem in multiple extended target tracking is formulated using the track-oriented probabilistic graphical model.Then,the belief propagation algorithm is used to solve the probabilistic graphical model,and a simplified measurement set is constructed to reduce computational complexity.The simulation results show that the proposed algorithm has advantages over the traditional data association methods in terms of accuracy and efficiency.2.The coupling between target association variables and the inaccuracy in extension state estimation of the track-oriented probabilistic graphical model have been studied.In the trackoriented probabilistic graphical model,all the extended target variable nodes are connected by factor nodes.The computational complexity is proportional to the square of the measurement cell number.This results in high computational complexity when the measurement cell number is large.Thus,it can only be applied in scenarios with few clutter measurements and high demand for computational efficiency.In addition,the consistency constraints of the track-oriented probabilistic graphical model only guarantee the consistency between extended target variables.Although the consistency constraints are satisfied,some extended targets may be associated with measurement cells with few measurements.In order to solve these problems,a data association algorithm for multiple extended target tracking is proposed based on a hybrid probabilistic graphical model.The proposed algorithm avoids the coupling between target association variables using measurement association variables.Furthermore,according to prediction information,the simplified measurement set is improved to obtain high accuracy in extension state estimation.The simulation results show that the proposed algorithm has advantages over the trackoriented probabilistic graphical model in terms of robustness,accuracy,and efficiency.3.Target spawning of the Poisson multi-Bernoulli(PMB)filter has been studied.The PMB filter performs worse when target spawning arises.This is due to the fact that the PMB filter treats spawned targets as birth targets,ignoring the surviving targets' information.In order to solve this problem,PMB filter with target spawning is proposed,in which the spawned targets are modeled by Poisson point processes.Then,the prediction and update of PMB filter with target spawning are formulated.Furthermore,the spawning model is applied to two PMB filters.The simulation results show that the PMB filters with target spawning have better performance when target spawning occurs.4.The inaccuracy in the marginal missed detection probabilities of measurement-oriented marginal Me MBer/Poisson(MOMB/P)filter has been studied.The marginal missed detection probabilities are inaccurate when the hypothesized tracks are separated,even if the measurement is close to the target state.This is due to the fact that the effect of measurement on predicted target states may be weakened,which can result in bias of cardinality estimation.In order to solve this problem,an improved MOMB/P filter is proposed.The proposed algorithm considers the measurement information in the missed detection hypotheses to obtain more accurate marginal association probabilities.The simulation results show that the improved MOMB/P filter has higher accuracy than the MOMB/P filter when targets are separated.