一种自适应色彩融合的Mean Shift跟踪算法*

针对现有的MeanShift算法使用单纯的颜色特征不能适应光线及背景的变化，易受颜色相近物体干扰的问题，提出了自适应色彩融合方法来提高跟踪性能。对背景以极坐标的形式进行不等间隔采样，以融合后的目标直方图与背景直方图具有最小相似性为原则搜索色调与饱和度的最佳线性融合系数；考虑背景与目标的渐变，跟踪过程中在最佳融合系数的自适应调整邻域内调整融合系数；能够有效处理相似物体和颜色相近的大背景带来的干扰。视频序列跟踪结果表明，提出的方法能够实时、稳定地进行跟踪。

Mean Shift tracking with adaptive combinational color features

Consider the issue of simple color features which were not robust within the Mean Shift framework, an adaptive tracking algorithm using combinational color features was proposed to improve tracking performance. The hypothesis was that the features best discriminate between object and background were also best for tracking the object. Several pixels were sampled in different intervals with polar coordinates in background. A best group of coefficients of hue and saturation was selected by minimizing the similarities between the sampled pixels of background and object. Consider the gradual changing appearance of both tracked object and scene background, a search in a coefficient neighborhood was performed to get the next most adaptive coefficients. Examples are presented that demonstrate that this real time algorithm can avoid confusion which caused by similar object and background in tracking.