Random Finite Set Based Multi-Target Tracking Algorithms With Noise Outliers

Multi-target tracking(MTT)is one of the most important research topics in the field of information fusion.It has been widely used in military and civilian fields such as air defense,traffic control,intelligent monitoring,and automatic driving.Due to the fact that the targets may appear and disappear at any time with a random pattern,the filter is designed to estimate not only the individual states but also the time-varying number of targets.Recently,multi-target tracking,based on random finite set(RFS)theory,has received intensive attention from both academic and military fields because of its unnecessary of data association.In the RFS based approaches,the multi-target state and multi-target measurement are modelled as RFSs,and the multi-target Bayesian filtering framework based on RFS is formulated by using Finite Set Statistics(FISST).Three popular approximations of the optimal multi-target Bayesian filter have been proposed,i.e.,Probability Hypothesis Density(PHD)filter,Cardinalized Probability Hypothesis Density(CPHD)filter and Cardinality Balanced Multi-Target Multi-Bernoulli(CBMeMBer)filter.These three aforementioned filters can achieve a good performance based on the assumption of Gaussian noise.Unfortunately,they may fail in some practical scenarios with outliers in the process and measurement noise which may not follow the Gaussian assumption.This dissertation studies multi-target tracking problems with noise outliers based on RFS theory.The main contributions of the dissertation are summarized as follows: 1.To solve the multi-target tracking problem with measurement noise outlier,a novel PHD filter named Gaussian inverse Wishart mixture PHD(GIWM-PHD)filter is proposed.In the proposed method,auxiliary variable is introduced to model the measurement noise.Then the multi-target tracking problem with measurement noise outlier can be transformed into the joint estimation problem of state and measurement noise covariance by approximating the multi-target intensity as a Gaussian inverse Wishart mixture form.Furthermore,the GIWM-PHD filter is derived via using variational bayesian technique to estimate the parameters of the Gaussian inverse Wishart mixture components simultaneously.The simulation results indicate that the proposed GIWM-PHD fliter can achieve a good performance for multi-target tracking problem with measurement nosie outlier.2.To solve the multi-target tracking problem with process and measurement noise outliers,a novel CBMeMBer filter called the Student's t distribution mixture CBMeMBer(STM-CBMeMBer)filter is proposed.In the proposed method,in order to catch the heavy tailed process and measurement noise with outliers,the Student's t distribution is used to modell the process and measurement nosie.Then the closed-form solution to the CBMeMBer recursion is derived by approximating the probability density parameter of the multi-Bernoulli as a Student's t distribution mixture.The proposed STM-CBMeMBer filter is a generalization of the existing Gaussian mixture CBMeMBer(GM-CBMeMBer)filter,and it will reduce to the GM-CBMeMBer filter as the degree of freedom approaches infinity.The simulation results indicate that the proposed STM-CBMeMBer fliter can deal with the multi-target tracking problem with process and measurement nosie outliers.3.In order to solve the performance degradation of Gaussian mixture CPHD(GM-CPHD)filter induced by process and measurement noise outliers,a novel CPHD filter based on Student's t distribution is proposed.The method introduces Student's distribution to model the heavy tailed process and measurement noise.By approximating the multi-target posterior intensity as a Student's t distribution mixture form,the linear(Student's T Distribution Mixture CPHD,STM-CPHD)and nonlinear(Robust Student's T Distribution Mixture CPHD,RSTM-CPHD)closed-form solution of the CPHD are derived respectively.Furthermore,the moment matching algorithm is used to prevent the infinite growth of the degree of freedom of student's t distribution.The existing GM-CPHD filter can be deemed as a special case of the proposed algorithm when the degree of freedom of the Student's t distribution tends to infinity.The simulation results demonstrate that the proposed filter can achieve a robust multi-target tracking performance in the presence of process and measurement noise outliers.4.Aiming at the problem of multi-target tracking performance degradation of the Gaussian Mixture Particle PHD(GMP-PHD)filter,the Gaussian Mixture Particle Flow PHD(GMPF-PHD)filter is proposed.This method recursively propagates the multi-target posterior intensity approximated by a Gaussian mixture form.Then the particles representing the Gaussian mixture components are updated by particle flow in order to approximate the posterior intensity more accurately.The proposed filter can achieve a better performance than the GMP-PHD filter due to making full use of measurement information.Moreover,due to the avoidance of clustering and resampling operations,the proposed algorithm can fix the erroneous state estimation caused by unstable clustering algorithm,and it has less computational complexity than that of the particle PHD filter.The simulation results demonstrate that the proposed GMPF-PHD filter can achieve a good performance for both linear and nonlinear problems.