Abstract:

In the single-target long-term tracking process, in order to avoid the effects of rapid movement and motion blur noise, and deal with the problem of tracking failure recovery mechanism, this paper presents a TLD (tracking-learning-detection) object tracking algorithm based on spatio-temporal context similarity (TLD-STCS). Firstly, conduct a detector study. Then, use the STC tracking algorithm the next frame, and calculate the approximation of the candidate target and the previous frame, including conservative similarity and motion similarity, to track the results of the effective judgments, if the decision is valid, the output process is the same as that of the TLD, if the judgment fails, the temporal space model at this time is added to the target space-time model. Next, calculate the confidence pattern for the multiple candidate target position detected by the detection module, output the detection target with the highest average confidence value, and update the target space-time model, a single cluster is detected for direct output. Finally, update the relevant parameters of classifier to improve the detection accuracy by online learning. Conducting experiments in the benchmark data set demonstrate the success of TLD-STCS. The results show that TLD-STCS algorithm can adapt to target tracking, rotation and other complex scenarios under the target tracking, has a high degree of robustness, especially in the case of fast moving and moving blur with good anti-interference ability and high success rate.

Key words: target tracking, spatio-temporal context, detection classifier, target space-time model, confidence map

摘要：

在单目标长期跟踪过程中，为了避免快速移动、运动模糊的噪声影响以及解决目标出视角后再出现的跟踪无法恢复问题，提出了一种基于时空上下文相似性的TLD（tracking-learning-detection）目标跟踪算法（TLD object tracking algorithm based on spatio-temporal context similarity，TLD-STCS）。首先进行检测分类器的学习。然后利用STC跟踪算法进行下一帧计算，对计算得到的获选目标与前一帧目标进行空间上下文的相似性计算，即保守相似度计算以及运动相似度计算，进行跟踪结果的有效判断，若判定有效，则输出过程与TLD的一样；如果判定失效，将此时的上下文时空模型加入到目标时空模型。对检测模块检测到的多个候选目标位置计算其置信图，输出平均置信值最大的检测目标，并对目标时空模型进行更新，如果检测到单聚类框就直接输出。最后进行在线学习来更新分类器的相关参数，改善检测精度。在不同测试视频序列上进行算法对比验证，结果表明，TLD-STCS算法能自适应目标遮挡、旋转等复杂情景下的目标跟踪，具有很高的鲁棒性，尤其是在目标快速移动且运动模糊情况下具有很好的抗干扰能力和很高的成功率。

关键词: 目标跟踪, 时空上下文, 检测分类器, 目标时空模型, 置信图