核函数自适应的Mean Shift目标跟踪算法

Algorithm of target tracking based on Mean Shift with kernel function adaptive

为解决Mean Shift算法无法对核函数带宽进行自适应更新的缺陷,提出目标质心的分布散度与多级正方形匹配结合的核函数带宽的更新算法。利用目标质心点的分布散度和增量试探法计算几个目标的可能变化尺度,采用多级正方形匹配计算各回字形区域间的Bhattacharyya距离预测目标的尺度变化趋势,对该趋势下的几个目标尺度进行Bhattacharyya距离对比,Bhattacharyya距离最大者为当前核函数的带宽,即目标的尺度。该策略减少了背景噪声的干扰以及每次计算目标收敛区域时的冗余像素的干扰。实验结果表明,该策略优于增量试探法和传统的核函数带宽不变化的方法,在时间代价上略低于增量试探法。

To solve the problem that Mean Shift algorithm itself is unable to update the adaptive bandwidth of kernel function , the combination of distribution of target centroid divergence and multi‐stage square matching based kernel function bandwidth up‐dating algorithm was put forward .Firstly the target centroid point divergence and incremental trials were utilized to calculate the distribution scale of several objectives ,then to predict change trend of the scales of the targets ,the Bhattacharyya distance w as calculated using multi‐stage square matching ,then comparing the scales of these targets with Bhattacharyya distance ,the maxi‐mum distance of Bhattacharyya was taken as the bandwidth of the current kernel function ,namely the scale of goals .The strate‐gy reduces the background noise interference and the interference from redundancy for each pixel when calculating the target area of convergence .Experimental results show that this strategy is superior to the incremental trials and the traditional kernel func‐tion method which can not change the bandwidth ,and the time cost is less than that of the incremental trials .