基于KCF相似度的TLD目标跟踪算法

研究持续单目标跟踪算法时,TLD的架构是一种值得借鉴的方式,但由于其本身的一些缺陷,当跟踪的目标在出现遮挡、快速移动以及光照变化等复杂情况下,TLD跟踪算法容易发生目标跟丢的情况,并造成误差的逐渐累积。鉴于TLD跟踪算法以中值流跟踪算法作为跟踪器存在局限性,本文提出一种基于KCF相似度的TLD目标跟踪算法(TLD-KCFS)。采用KCF算法对TLD跟踪进行实时监督,通过跟踪结果计算出相似度,利用相似度进行检测模块切换的判断,并结合两种结果调整目标跟踪框。通过对多类型的视频序列进行测试后可知,TLD-KCFS算法在遇到模糊和快速移动、目标遮挡、光照变化等复杂情况时,可以稳定且良好地输出跟踪结果,鲁棒性较好,适用于长时间目标跟踪。

A TLD object tracking algorithm based on KCF similarity

The architecture of the tracking learning detection (TLD) algorithm is a good reference for studying long time single object tracking algorithms. However, due to some defects of its own, the TLD algorithm tends to cause the accumulation of errors and the occurrence of losing objects in complex situations such as fast moving, occlusion and light changes. Given the limitation of the median flow algorithm as the tracker in the tracking module of the TLD algorithm, we propose a TLD object tracking algorithm based on KCF similarity (TLD-KCFS). The KCF algorithm is used to monitor the TLD tracking in real time. The similarity is calculated via the tracking results to judge the switching of the detection module, and the bounding box is adjusted by the combination of the two results. Tests on several different types of videos show that the TLD-KCFS algorithm can achieve stable and good tracking output in complex situations such as blur, fast moving, occlusion, and light changes. It is robust and suitable for long time object tracking.