基于Mean-Shift优化的TLD视频长时间跟踪算法

Long-term video tracking algorithm of optimized TLD based on Mean-Shift

针对TLD(tracking learning detection)算法同时包含了跟踪、检测和学习三个部分,具有较高计算量的缺点,提出了采用Mean-Shift算法替换原TLD跟踪器部分的光流跟踪算法。该优化方法利用具有计算量小的MeanShift算法替换计算量较大的光流法进行跟踪,以通过目标模型和候选目标模型之间的巴氏系数与阈值的比较来判定跟踪失败的自检测,并通过计算Mean-Shift跟踪返回的目标框和上一帧TLD返回的目标框之间的相似度来进一步得到跟踪的有效性,在发生跟踪失败时由检测器重新初始化跟踪。实验结果表明,该优化方法在视频长时间跟踪算法中具有较高的鲁棒性和准确性,并且与原TLD算法相比,该优化方法在跟踪速度上得到了提升。

TLD algorithm combines tracking,learning and detection simultaneously, so its computation is high. This paper adopted Mean-Shift algorithm to substitute the optical flow tracker of TLD. Considering optical flow tracking has high computa- tion,this optimized algorithm made use of less computation of Mean-Shif! to replace optical flow in original TLD. Compared the Bhattacharyya coefficient between target model and target candidate model with the given threshold,it decided whether track failed. Measured similarity between the bounding box returned by Mean-Shift anti the bounding box returned by TLD last time, it decided the confidence of the track. If it was failure, the detector would re-initialize the tracker. The experiments show that the optimized algorithm can acquire high robustness and accuracy in long term tracking in video, and it obtains a more rapid tracking speed than original TLD algorithm.