Research On Methods Of Target Tracking Based On Particle Filter And Box Particle Filter

Target tracking problem has extensive application prospects in in the field of military and civilly. However, the false alarm, clutter, the randomness of target number, missing detection, nonlinear measurement and other uncertainty factors have made it become the research hotspots and difficulties in the field of academia and engineering field. The focus of the early research work was on solving the association decision problem between measurements and true targets. People have proposed a series of free data associating target tracking algorithms based on random finite set along with the proposed finite set statistics theory. Probability hypothesis density filter is the most representative multi-target tracking algorithm which is meant to work in the environment with unknown number of targets.Particle filter is a filtering method based on Bayesian recursive equations and sequential Monte Carlo simulation. Particle filter has better universality and scalability of algorithms compare to Kalman filter and other algorithms, so it is widely used to deal with non-linear and non-Gaussian filtering problems. This method utilizes discrete random measure composed of samples and related weights to describe the posterior probability distribution of true state, and get the state estimation of the latest moment through predication and update steps. Due to the most target tracking problems in practical are nonlinear, so particle filter is one of the basic tools to solve the target tracking problem, which is complex state estimation problem. Box particle filter is an improved algorithm which combines interval analysis and particle filter, it extends the properties of particle and reduce the computational complexity of particle filter. Box particle filter can achieve state estimation of high precision with less particle number compare to traditional particle filter.The target tracking algorithms which based on particle filter is the research focus of this thesis. The research content of this thesis is as following:1) The impacts of resampling algorithms on multitarget particle filter. This thesis first theoretically analyzes three commonly used resampling algorithms(Multinomial Resampling, Stratified Resampling, Systematic Resampling), which are proposed in order to solve the particle degradation in particle filter, and mainly compares their filtering precisions and computing time in the settled multi-target tracking scenario by combining them with free clustering particle probability hypothesis density filter. Simulation shows that the computational complexity of multinomial resampling is higher than that of the other two resampling algorithms under different particles sizes, while the filtering precisions of stratified resampling and systematic resampling are better than multinomial resampling, the computation efficiency of multinomial resampling is relatively lowest. Other than that, stratified resampling and systematic resampling share similar computational complexity, the advantage of systematic resampling becomes distinct along with the increase of particle number, its computation efficiency is relatively highest when the particle number is large.2) For the potential existing curse of dimensionality in the implenmemtation of sequential monte carlo PHD filter, this thesis presents marginalized free clustering particle PHD filter which marginalizes the state space on the foundation of the existing free clustering SMC-PHD filter. The presented algorithm utilizes the thoughts marginalized particle filter, to decomposable state space model(linear/non-linear), the algorithm adopts sequential monte carlo method and linear filter to predict and update the nonlinear and linear components in the PHD recursion formula by utilizing the principle of marginalized particle filter. At the same time, the algorithm extracts multi-target state in the update step, thus it avoids the problem caused by clustering operation in SMC-PHD filter, and the clustering operation is meant to extract multi-target state. The simulation results show that the presented algorithm reduces the system dimension and improve the overall performance of multi-target tracking compare to SMC-PHD and existing FCP-PHD.3) At last, this thesis studies the single target tracking problem of Box-Bernoulli filter under two typical nonlinear cases. Firstly, the motion model of target and measurement model are established based on Bernoulli random finite set, then it mainly focuses on the method to solve constraint satisfaction problems of two typical nonlinear tracking(Range-Bearing tracking, Range-Bearing tracking with Doppler measurement) with constraints propagation algorithm combined with Box-Bernoulli filter. Finally, it compares the performance of single target tracking of SMC-Bernoulli filter and Box-Bernoulli filter in the two different nonlinear cases through experiment. The simulation verifies the tracking performance of Box-Bernoulli algorithm in the two nonlinear conditions.