Research On Multi-target Tracking Algorithm Based On Random Finite Set Modeling

Multi-target tracking algorithms have a wide range of applications in aerospace,computer vision,monitoring,and automatic control areas.For more than 10 years,the multi-target tracking algorithms based on Finite Set Statistical Theory(FISST),which models multi-target as Random Finite Set(RFS),has received extensive attention and research applications because of its “engineering friendly” nature.However,FISST algorithms are mostly derived in the context of known prior information,such as known clutter models.In addition,according to the different multi-target density modeling of the filter,iterative derivation for its multi-sensor background is also fundamentally different.The paper condiders conjugate density modeling method.In response to the above problems,this paper first derives an adaptive clutter rate PMBM iterative structure,based on which an adaptive clutter rate Poisson Multi-Bernoulli(PMB)filter is implemented.Then,using the idea of two-step multi-sensor measurement space partition,a centralized multi-sensor PMBM iterative structure is designed,based on which the multi-sensor PMB filter is implemented.Finally,a new structured of Multi-Target Multi-Bernoulli(Me MBer)filter based on Bayesian closed structure derivation is implemented,and a new two-step partitioning algorithm is designed on this basis.Then,the single sensor algorithm is extended to the centralized multi-sensor application background.The specific research contents are summarized as follows:Firstly,the multi-target tracking algorithm needs to input a lot of priori informations when it is applied.When the algorithm is applied in some environment with unknown prior information,the multi-target tracker needs to be designed as an adaptive structure.Based on modelling the unknown clutter rate as Gamma distribution,this paper derives an adaptive clutter rate PMBM iterative formula with the ability to adapt to the unknown clutter rate.For the specific implementation of iterative,it is designed into a data structure similar to the Multi-Hypothesis Tracking(MHT)algorithm,and the Gibbs sampling algorithm is used to solve the rank assignment problem involved in the implementation process.After each update step,the density of Multi-Bernoulli Mixture(MBM)is approximated to the density of Multi-Bernoulli(MB),and finally the Adaptive Clutter Rate Poisson Multi-Bernoulli(A-PMB)filter is obtained.The simulation results verify that the proposed algorithm can perform real-time joint estimation of both unknown clutter rate and multi-target state information.Secondly,based on the two-step multi-sensor measurement partitioning algorithm,a multi-sensor PMBM iteration is derived.The key of the derivation lies in the design of the multi-sensor measurement likelihood function adapted to the multi-sensor measurement space partitioning algorithm,so as to re-derive the PMBM iteratively based on Probability Generating Functional(PGFl)and functional differential theory.Aiming at the problem of traditional multi-sensor measurement space partitioning algorithm that cannot effectively perform multi-target measurement space partitioning for multiple closely spaced targets,an improved method is proposed.Simulation verifies that the proposed improved method can make the filter track multiple closely spaced targets stably.Thirdly,in order to solve the problem that the Me MBer filter cannot be applied to challenging tracking environments due to two strong approximations,a novel structure of Me MBer filter based on the conjugate distribution is proposed.Simulation verifies that the proposed algorithm can be applied to more challenging tracking environments and has better tracking performance than Labeled Multi-Bernoulli(LMB)filter.Finally,a new two-step multi-sensor measurement space partition algorithm based on marginal probability distribution calculation is proposed,based on which the novel structure of Me MBer filter is extended to its multi-sensor form.Simulation verifies that the proposed algorithm can not only track multiple closely spaced targets,but also has better calculation efficiency than that of traditional multi-sensor Me MBer filter,which used the traditional two-step partitioning algorithm.