Improvement Particle Filtering Algorithm for Nonlinear Non-Gaussian models

An improved particle filter algorithm is proposed for the highly non-linear passive location and tracking system in which the common tracking filters often faile to catch and keep tracking of the emitter. Firstly, the algorithm uses limite gaussian mixture model to approximate the posterior density of states. Secondly, in order to solve the problem in which the stochastic observation noise influence the accuracy of particle weights, an improved based on averaging likelihood functions with diverse proportion is proposed. According to χ2 testing, the new method combines multi-observations to compute the likelihood functions of each particle and then average them with single observation computes the likelihood functions of each particle to update particle weights. The method not only reduces the influence of the stochastic observation noise to particle weight but also improves real-time. Finally, the traditional process of particle filter resampling is replaced by the aitken-deterministic annealing expectation maximization (A-DAEM) algorithm, avoiding to a local maximum and reducing the effects cause by sampling depletion. Simulation results show that the algorithm outperforms the one based on PF-ALDP and the other based on EMGMPF in tracking accuracy and stability. Therefore it is more suitable to the nonlinear state estimation.