Research And Improvement Of Resampling Algorithm In Particle Filter

Particle filter is an algorithm based on Monte Carlo and recursive Bayesian estimation. Its basic idea is to use series of weighted samples to approximate posterior probability density distribution in the state space and to present integral operation by sample mean value. In principle, particle filter can realize any state estimation, especially shows excellent performance in non-linear/ non-Gaussian problems while Kalman filter and extended Kalman filter lose efficiency. Particle filter has been widely used in the fields such as target tracking, vision tracking, fault diagnosis, navigation position and radio communication etc. Resampling algorithm is an important step of particle filter, and is one of important methods to solve degeneracy problem of the particle filter.This thesis firstly studies particle filtering algorithm in depth, introduces realization theory and steps of the algorithm in detail, and briefly discusses choices of importance density function and resampling algorithm; then this thesis theoretically analyzes four classic resampling algorithms including multinomial resampling, residual resampling, stratified resampling and systematic resampling, from points of resampling realization principle, resampling quality and computation complexity, and compares performances of the four resampling algorithms via experiment simulation.Based on residual resampling and residual systematic resampling, this thesis presents an improved residual resampling algorithm, the most difference of the improved algorithm compared with residual resampling algorithm lies in that this improved algorithm does not compute the particle copied times separately, but firstly computes the sum of the product of particle weights and the total particle number, then rounds, ensuring the invariability of the particle number before and after resampling, that is to output particles after gaining the accumulation of particle copied times, so that this improved algorithm can avoid resampling operation for the residual particles existing in the residual resampling, reduce computational complexity, and improve the running efficiency to some degree. On the basis of the predecessors' successfully introducing evolution scheme and genetic algorithm into particle filter, this thesis introduces Gaussian particle swarm optimization, optimizes the particles before resampling so as to make particles intelligently cooperate to move towards the posterior probability density distribution of the true states. Simulation results show that the estimation performance of this algorithm is superior to that of traditional particle filter and particle filter with the introduction of MCMC (Markov Chain Monte Carlo).