一种抑制背景干扰的粒子滤波人脸跟踪算法

首先从理论上推导出基于MMSE状态估计在背景干扰下的偏差,然后通过实验说明基于MAP状态估计存在的问题,在此基础上提出一种新的跟踪目标状态参数估计方法.即在视频目标跟踪过程中,按照粒子权值大小的准则,筛选适当数量具有较大权值的粒子进行目标状态估计.由于该算法利用了参考目标与候选目标相似度大的特性,所以可以有效地剔除背景以及伪目标的影响.实验结果表明,该算法具有很好的鲁棒性,并且提高了在背景干扰下目标跟踪的精度.

Tracking algorithm based on particle filter with restrained background disturbance

The estimation discrepancy of the minimum mean square error(MMSE) under background disturbance is theoretically proved,and the simulation experiments demonstrate that the same problem exists in the maximum a posteriori(MAP).Based on these,a novel estimation method of tracked target status is proposed,which selects some particles with large weights for estimating the tracked target state parameters.By using the high similarity of the reference model and the candidate one,the possible influence of the background or the fake target can be efficiently eliminated.Experimental results show that the new tracker has excellent robustness and high tracking accuracy under the background disturbance.