Labeled Box-particle Filter And Its Implementation For Multiple Extened/Group Target Tracking

The box-particle filter(BPF)is a generalized particle filter which results from the synergy between interval analysis and Monte Carlo method.Using interval analysis,particles can be represented as controllable multidimensional rectangular boxes having nonzero volumes in the state space,which is so called box-particles.Key advantages of BPF against the standard particle filter(PF)are its reduced computational complexity and its suitability for imprecise measurements.On the other hand,random finite sets(RFSs)theory based filter is the trend for the target tracking field,since its correspondence at set level can effectively avoid complex data association,therefore reduce computational complexity.Combining with BPF and RFSs filter,box-particle filter based on random finite sets has high computational efficiency and can avoid data association which heen widely using in the field of multiple target,extended target and group target tracking.However,these algrithms can not distinguish different targets,since the elements in a random finite set is in disorder,the filter cannot recognize the correspondence between the current state and previous state of a random finite set and therefore cannot identify different target trajectories.It is more like filtering rather than tracking.To address this problem,labeled box-particle filter(LBPF)under the framework of traditional BPF and RFSs filter has been proposed,and the multi-extended / group target tracking algorithm based on labeled box particle random set filtering in complex scenes has been also studied and implemented in this thesis.The main work includes:(1)Research on LBPF algorithms.Aiming at the problem that the traditional BPF cannot distinguish different targets,a labeled box particle filter is proposed.It can not only inherit the advantages of BPF which reducing computational complexity and dealing with imprecise measurements,but also implement the targets discrimination and achieve different trajectories.The combination of LBPF and RFSs can avoid data association while achieving trajectories discrimination and management.(2)Research on LBPF for multiple extended targets tracking based on RFSs.Aiming at the problem of box particle information degradation,an information supplement step is proposed to expand the volume of box particles after resampling,so that the information contained in the box-particles can be back to the initial time,which can effectively enhance the stability of LBPF,and significantly improve the performance of LBP based on RFSs filters for target tracking in complex scenarios.Aiming at the problem that multiple extended tracking algorithm based on RFSs cannot distinguish different targets and PHD filter achieve less accurate number estimate than CPHD filter,the LBP implementation is given for multiple extended target tracking based on CPHD filter.After partitioning the measurements belonging to the same extended target into the same measuring cell,the LBPF adds the same label to the box particles divided into the same measuring cell,and the particles resampled from different measuring cells are added with different labels;therefore the whole disordered box particles set are divided into several label box particles subsets by labels,which can both realize trajectories management and targets discrimination according to the inherited labels of state estimates.The simulation results show that ET-LBP-CPHD filter can effectively distinguish and identify different extended targets without changing the CPHD filtering framework,and obtain different extended target trajectories.(3)Research on LBPF for multiple group targets tracking based on RFSs.According to the radar resolution unit,group target tracking can be classified into resolvable group target tracking and unresolvable group target tracking and the LBP implementations for these two kinds of group target tracking based on PHD filter are both presented in this thesis.Simulation results show that compared with GT-BP-PHD filtering,GT-LBP-PHD filtering can not only distinguish and different targets and achieve different trajectories,but also effectively avoid the instability of clustering in state estimation step,which accommodate GT-LBP-PHD filtering to more complex scenarios.