Mean Shift跟踪的背景优化及动态分析

Background Optimization and Dynamic Analysis of Mean Shift Tracking

针对传统Mean Shift跟踪算法在进行目标跟踪时背景带来的定位偏差及由于缺乏相应的跟踪状态分析策略而易陷入局部最小值的缺陷,提出了两方面的改进措施。一是将跟踪窗口内的目标和背景区分开来,对背景像素定义新的特征模型以弱化背景像素对目标模型的影响。二是将跟踪窗口进行分块处理,综合考虑每个子块相似度的大小变化建立判断准则,对跟踪状态进行动态实时分析,以判断目标是否存在遮挡:如部分遮挡,则应用没有被遮挡的子块位置偏差对目标进行定位;如完全遮挡,则采取相应的二维线性预测方案根据先验信息对目标进行定位跟踪。将该方法应用于人物跟踪中进行实验,实验结果表明,该方法有效改善了Mean Shift跟踪算法的不足,对于复杂条件下的运动目标跟踪具有很好的鲁棒性。

Considering the defects of traditional Mean Shift when used in target tracking,i.e.,the locating error caused by background and local optimum due to no corresponding status analysis strategy,two kinds of improved methods were proposed.One is to differentiate the target from the background in tracking window,and create new feature model to background pixels in order to weaken the influence of background on target model.The other is to divide the tracking window into a number of fragments.Considering comparability of each fragment entirely,the judgment rule is created to make dynamic real-time analysis to tracking status,then to judge whether there is shelter or not.If the target is sheltered partially,the location windage of fragment which is not sheltered should be used for locating.When the target is sheltered entirely,the two-dimensional linear polynomial method can be used according to the priori information.The improved method was used for human tracking in experiment,and the result indicated that the improved algorithm can overcome the defects of the traditional Mean Shift,and has fine robustness for moving target tracking under complex condition.