Mean Shift跟踪算法中尺度自适应策略的研究

标准Mean Shift跟踪算法缺乏尺度自适应机制,一种常见的尺度自适应策略是在上一帧尺度及上一帧尺度基础上增/减10%3个尺度下执行3次标准Mean Shift算法来确定本帧的尺度。本文在一组典型场景下对这种方法进行了实验研究,发现它存在两个缺陷,即有时不能防止尺度在小于真实尺度处徘徊;对快速尺度变化适应性差。其中任何一个缺陷都可能引起大的尺度定位偏差,从而降低跟踪器的鲁棒性。在对上述缺陷深入分析的基础上,修正了最优带宽的判别条件,给出了自适应滤波器参数的设计方法,从而得到了一种改进的尺度自适应算法。多种场景下的实验结果表明了这种算法的有效性。

Researches on Scale Adaptation Strategy in Mean Shift Tracking Algorithm

The standard Mean Shift tracking algorithm is lack of scale adaptation mechanism.A familiar scale adaptation strategy is to determine the scale in current frame by running the standard Mean Shift tracking algorithm three times respectively based on the previous scale,its 10% up and 10% off scales.In this paper,the algorithm was tested on numerous typical scenes and its two drawbacks are found:(1) sometimes it may be stuck in the scale smaller than the real scale;(2) it often responds poorly to rapid scale changes.Such drawbacks can introduce additional scale error and thus increase the risk of missin tracking.Through analyzing above drawbacks in detail,we propose the revised scale adaptation algorithm,in which the criterion of optimal bandwidth selection is modified and adaptive filtering parameter is introduced.Experiment results in numerous scenes show the effectiveness and efficiency of the improved algorithm.