Research On Multi-target Tracking Methods Based On Random Finite Set

With the rapid development of modern radar,laser,infrared,sonar and other sensor technologies,multi-target tracking(MTT)has become one of the important research contents in the field of information fusion,which is widely used in air traffic control,missile defense,intelligent monitoring,vehicle tracking and other military and civil fields.The purpose of multi-target tracking is to estimate the unknown and time-varying number and states of the targets from the unknown source measurements received by the sensor under the complex background of the target that may be newborn,dead,missed detection and the presence of clutter and noise interference.Besides,with the application of high-resolution sensors,the target generates multiple measurements,which brings the estimation problem of extended states such as measurement number,size,shape and direction,and the more complex problem of multiple extended target tracking.The multi-target tracking algorithms based on data association have the limitation that the computational complexity increases with the increment of the target number,which cannot solve many complex multi-target tracking problems.In recent years,the multi-target tracking algorithms based on the random finite set(RFS)theory can avoid data association,use the rigorous and elegant finite set statistical theory to describe the multi-target tracking problem,and iteratively propagate the multi-target posterior density,which have been widely concerned and deeply studied.Based on the random finite set theory and the Bayesian filtering framework,this thesis focuses on radar multi-target tracking and studies the problem of multi-point target tracking under the condition of unknown process and measurement noise covariances,and the key problems of measurement number modeling,unknown detection probability estimation and the size estimation of the non-ellipsoidal extended target in multiple extended target tracking.The main contents and research results are as follows:1.To solve the problem of unknown covariances of process and measurement noise in the traditional Poisson multi-Bernoulli mixture algorithm,a robust Poisson multi-Bernoulli mixture algorithm based on Gaussian inverse Wishart inverse Wishart is proposed.To jointly estimate the kinematic state,prediction state covariance,and measurement noise covariance,the augmented state is modeled as Gaussian inverse Wishart inverse Wishart(GIWIW).By introducing the GIWIW model into the Poisson multi-Bernoulli mixture(PMBM),and using the variational Bayesian method to approximate the multi-target posterior density as GIWIW form,the closed form of GIWIW-PMBM is obtained.The simulation results show that the proposed algorithm is the best compromise filter in terms of computational complexity and tracking performance,compared with the standard PMBM and GIWIW generalized labeled multi-Bernoulli(GIWIW-GLMB).2.To solve the problem that the traditional Poisson extended target probability hypothesis density cannot accurately estimate the number of spatially close extended targets,an extended target tracking method based on binomial extended target probability hypothesis density is proposed.It is assumed that the measurement number of the extended target obeys binomial distribution,which is introduced into the filtering framework of probability hypothesis density.The binomial extended target probability hypothesis density is derived by the finite set statistical tools,and the Gaussian mixture implementation is given.Simulation results show that,compared with the traditional Poisson probability hypothesis density,the proposed algorithm can achieve better tracking performance when the extended targets are close to each other.3.To solve the problem of unknown detection probability in binomial extended target probability hypothesis density algorithm,an extended target tracking algorithm based on beta Gaussian probability hypothesis density is proposed.First,it is assumed that the detection probability obeys the beta distribution,and a Bayesian method is given to iteratively estimate the detection probability.Then the beta Gaussian model is established for the augmented state including the detection probability and the kinematic state.By introducing the model into the binomial extended target probability hypothesis density algorithm,the beta Gaussian probability hypothesis density filtering is derived.The simulation results show that the proposed algorithm can accurately estimate the detection probability and achieve better tracking performance than the gamma Gaussian extended target probability hypothesis density.4.To solve the problem that the size of the non-ellipsoidal target can not be accurately estimated by the traditional Gauss inverse Wishart probability hypothesis density filter,an extended target tracking algorithm based on improved Gaussian inverse probability hypothesis density is proposed.It is assumed that the shape of the extended target is an ellipse in the traditional Gaussian inverse Wishart probability hypothesis density algorithm,so it can not accurately estimate the extended state of the non-ellipsoidal target.The improved algorithm takes the non-ellipsoidal target as a reference ellipse with the same target size,designs the scattering matrix of the non-ellipsoidal target,and obtains the improved random matrix method.Then,the improved random matrix method is combined with the binomial extended probability hypothesis density filter.And the improved Gaussian inverse Wishart probability hypothesis density filter is obtained.The simulation results show that,compared with the traditional Gaussian inverse Wishart probability hypothesis density filter,the improved algorithm can achieve better performance when tracking the non-ellipsoidal extended targets that are spatially close.