Abstract: This paper proposes Mean Shift algorithm based on multi-scale feature extraction for fulfilling the target tracking in complex environment such as images with low contrast and to many similar targets. After the feature points being matched, next frame feature points are gotten. The center of next frame feature points is took as the center of searching window by which Mean Shift searching windows are continually modified and iteration deviation is reduced. Experimental resutls show that the robustness, precision and real-time performance of the algorithm are improved, and its iteration frequency is reduced.

Key words: multi-scale feature, feature extraction, feature point matching, Mean Shift, target tracking

摘要： 为在图像对比度较低、相似目标过多等情况下较好地实现目标跟踪，提出一种基于多尺度特征提取的均值漂移跟踪算法。前一帧目标区域的特征点经匹配得到后续帧目标区域的特征点，利用所得特征点集的中心坐标修正均值漂移搜索窗位置，以此为约束条件，减小均值漂移迭代产生的偏差。实验结果表明，该算法可以提高跟踪精度、鲁棒性及实时性。

关键词: 多尺度特征, 特征提取, 特征点匹配, 均值漂移, 目标跟踪

CLC Number: 

• TP301.6