基于特征区分度和区域生长的Mean Shift跟踪算法

Mean Shift Algorithm for Visual Tracking Based on Feature Discrimination and Region Growing

复杂环境下的目标跟踪是一个具有较多难点的任务。例如杂波干扰、严重遮挡、相似背景、运动不连续、光照变化等,都会给跟踪带来很大困难。针对上述问题,利用目标与背景的区分度,选取目标中独特特征建立模板,并对候选目标过滤背景特征,从而提高了目标在复杂背景下的匹配精度。同时为了解决Mean Shift算法搜索到的目标位置与真正目标存在偏差的问题,以目标中高区分度的像素点为种子点,进行区域生长,来获得准确的目标位置,并以此确定目标框大小。实验结果表明,文中算法在复杂环境下具有较好的跟踪精度和实时性能。

Object tracking under complex environment is a tough task. Some critical problems, such as the dis-turbance of the clutter, terrible occlusion, similar background, discontinuous motion and the changes of the illumi-nation will seriously influence tracking accuracy. Aiming at these problems, a method to inhibit background fea-tures is proposed, by selecting unique features of object with the discrimination between object and background tocreate object template, and also weighting features of candidates with the discrimination, and the matching accura-cy of the object in complex background is improved. Meanwhile, to eliminate the error between the object locationcalculated by mean shift algorithm and the actual object location, pixels with high discrimination value are selectedto be seed points for growing a precise object region, which is used to modify the location and the size of the object.Experimental results on challenging videos in complex environment show that the proposed algorithm has bettertracking accuracy and real-time characteristic.