基于MeanShift的跟踪学习检测目标跟踪改进算法

An Improved TLD Target Tracking Algorithm Based on Mean Shift

研究了一种将MeanShift和TLD结合的目标跟踪改进算法。当TLD跟踪框有较高的可信度时,以TLD输出的目标中心位置作为MeanShift跟踪算法的迭代起点;当置信度低时,将前1帧中目标跟踪框的中心位置作为跟踪的迭代开始点。比较表明,当目标被遮挡和抖动时,改进算法能实现稳定跟踪,实现了跟踪的鲁棒性。针对TLD算法通过均匀采样获得的特征点中存在较多的无用点,在TLD跟踪模块引入了更具鲁棒性的Susan角点作为目标的特征点。选择角点后采用金字塔LK光流法跟踪,跟踪过程中保留信息丰富的特征点,抑制了目标关键信息点集较少导致的跟踪漂移。实验表明,本文算法具备比较高的鲁棒性和实时性。

The TLD algorithm is sensitive to illumination change and clutter results in drift even missing,and the corresponding tracker designed based on the pyramid Lucaks-Kanade optical flow method needs vast computation. To overcome these shortcomings,an improved target tracking scheme by integrating mean-shift and TLD algorithm is proposed. When the confidence level of the TLD tracking box is high,the center position of target of the TLD output is used as the starting point of the mean shift tracking algorithm. When the confidence level is low,the center position of the target box in the previous frame is used as the iterative starting point for the mean shift. The results show that the improved algorithm achieves higher precision,especially for occlusion and target jitter. In order to solve the problem that there are more useless points in the feature points obtained by uniform sampling of TLD algorithm,a more robust Susan corner point is introduced into the TLD tracking module. This algorithm can track the object through the pyramid LK optical flow after selecting the corner. It not only preserves feature points with rich information during the tracking process,but also suppresses the tracking drift caused by more useless points. The results show that this method has high robustness and real-time resolution compared with the original TLD algorithm.