基于改进的MeanShift鲁棒跟踪算法*

徐海明1，黄山1,2，李云彤1

(1.四川大学电气信息学院，四川 成都 610065;2.四川大学计算机学院，四川 成都 610065)

摘要：Mean Shift跟踪算法在目标尺度变化大和被遮挡时存在较大的缺陷。针对这一问题，提出了一种基于多级正方形匹配的自适应带宽选择和分块抗遮挡的目标跟踪算法。该算法采用目标中心点的离散程度和增量试探法计算出可能的变化尺度，然后采用多级正方形匹配法预测目标的运动趋势，将巴氏系数最大者的尺度作为Mean Shift核函数新的带宽。同时，对前景目标进行分块，根据子块的遮挡程度自适应改变子块权重并按一定准则融合有效子块的跟踪结果。实验结果表明，该算法具有很好的鲁棒性。

关键词：Mean Shift；目标跟踪；多级正方形匹配；分块

中图分类号：TP391.41 文献标志码：A

doi:10.3969/j.issn.1007-130X.2015.06.019

收稿日期：*2014-07-14;修回日期：2014-09-24

作者简介:

通信地址：610065 四川省成都市四川大学望江校区基础教学楼B217

Address:Room B217,Basic Teaching Building,Wangjiang Campus,Sichuan University,Chengdu 610065,Sichuan,P.R.China

ArobusttrackingalgorithmbasedonimprovedMeanShift

XUHai-ming1,HUANG Shan1,2,LIYun-tong1

(1.CollegeofElectricalEngineeringandInformation,SichuanUniversity,Chengdu610065;

2.CollegeofComputerScience,SichuanUniversity,Chengdu610065,China)

Abstract:The Mean Shift algorithm has a defect in handling moving targets with large scale change or being obscured. In order to solve this problem, we propose a bandwidth-adaptive and anti-blocking tracking algorithm based on multi-level square matching and fragment. The proposed algorithm uses the centroid deviation of the target model and the bandwidth trials method to compute the possible scales. The motion trend of the target is predicted through the multi-level square matching method, and the scale of the largest Bhattacharyya distance of the candidate targets is selected as the new bandwidth of the Mean Shift kernel function. At the same time, we divide the target into several fragments, adaptively change their weights according to the degree of being obscured, and then fuse the results of effective fragments under certain rules. Experimental results show that this algorithm has good robustness performance on tracking targets.

Keywords:MeanShift;objecttracking;multi-levelsquarematching;fragment

1引言

基于MeanShift的跟踪算法本质上是一种梯度下降方法，对该算法进行迭代直至收敛到相似性函数的局部极大值。ComaniciuD等人[1]提出的MeanShift算法就是利用目标的颜色直方图和MeanShift向量寻找到最相似的团块区域，以此作为目标在当前帧中的真实位置。

现阶段在MeanShift算法方面主要研究了以下几个问题：(1)原始MeanShift算法在跟踪运动目标的过程中带宽是固定不变的，因此不适合用于跟踪目标有尺度变化的情况；(2)当运动前景目标和背景的颜色比较相似时，MeanShift算法由于过度依赖于运动目标的颜色信息，容易导致跟踪丢失；……