一种改进的粒子滤波算法及其在说话人跟踪中的应用

Improved Particle Filter Algorithm with Application to Speaker Tracking

针对非线性、非高斯系统状态的在线估计问题,提出一种改进的粒子滤波算法,该算法综合考虑"优选建议分布函数"和"重采样"两种并行改进滤波性能的方法.首先通过Unscented卡尔曼滤波器产生系统的状态估计,并在协方差预测阶段引入衰减记忆因子,消弱滤波器对历史信息的依赖,增强当前量测信息对滤波器的修正作用,从而产生一个优选的建议分布函数,有效抑制了粒子退化现象;接着在重采样阶段引入MCMC(Markov Chain Monte Carlo)方法来构造马尔科夫链产生服从目标分布的粒子,使样本更加多样化,有效避免了粒子枯竭问题.最后,通过系统仿真及说话人跟踪实验,证明了该算法的有效性.

An improved particle filter based on ＂optimum proposed distribution function＂ and ＂resampling＂ two parallel improving filtering methods is proposed for the on-line estimation problem of non-Gauss nonlinear systems.In order to weaken the effect of historical information and enhance the effect of up-to-date measurement,we introduce attenuation memory factor for generating the important density function based on the Unscented Kalman Filter（UKF） for a better performance in inhibiting the particle degradation problems in the new algorithm.The new particle filter can solve the problem of sample impoverishment by introducing Markov Chain Monte Carlo into rsampling stage to construct a Markov Chain and produce particles with target distribution.Finally,system simulation and speaker tracking experimental results show that the proposed new partical filter is superior to the standard particle filter on filtering performance.