基于重要性排序蒙特卡洛粒子滤波的物体跟踪算法

研究复杂背景下的物体跟踪方法. 提出一种用于物体跟踪的重要性排序马氏链蒙特卡洛粒子滤波算法. 算法利用少量加权初始粒子得到后验概率分布的初步估计,并通过重要性排序马氏链蒙特卡洛采样技术从该初步估计抽取新的粒子,以构建对应不同模态的多条独立马氏链,从而充分逼近真实后验概率分布的多模态. 所提出的算法自适应地根据当前模态分布构建多条独立马氏链,因此能够在多模态的复杂场景下准确估计目标状态的后验概率分布;同时,在构建马氏链的过程中,算法采用重要性排序策略确定历史样本被选为状态转移核的似然度,提高了小权重样本被选中的可能性,降低了在马氏链构建过程中陷入局部最优的概率. 仿真实验以及真实视频上所进行的实验显示,所提出的方法能够实现准确稳定的物体跟踪,且效果优于标准粒子滤波算法以及马氏链蒙特卡洛粒子滤波算法. 

Object Tracking Arithmetic Based on Importance Ordering Monte Carlo Particle Filtering

In order to obtain efficient object tracking under cluttered scenes, a method was proposed based on importance ordering Markov Chain Monte Carlo (MCMC) particle filtering. Firstly, a few authorized initial particles were made use to approximate the true posterior particle distribution. And then the new particles were drawn from the rough approximation with the proposed importance ordering MCMC sampling strategy to build several independent Markov Chains, corresponding one-to-one to the modes of the true posterior distribution, so as to approximate the multimode of the true posterior distribution. According to the current mode distribution, the method could establish several adaptive independent Markov Chains to approximate the posterior distribution of objects under multimode cluttered scenes. An importance ordering strategy was taken to make fully use of the history samples for state transfer decision, to increase the possibility that small weight samples could be selected and to decrease the probability that build process of Markov Chains got in local optimized. Simulation and verity experiment show that the proposed method can achieve stably and exactly object tracking, its performance is better than the standard particle filtering method and the MCMC particle filtering method.