基于Struck算法的遮挡目标跟踪

Struck算法是近年来综合性能较优的视觉目标跟踪算法,但对于较大比例遮挡或全部遮挡情形,算法性能下降明显.通过对Struck算法的分析发现,当遮挡出现时算法分类器会引入错误信息,从而导致跟踪失败或者跟踪漂移.在Struck算法框架的基础上加入遮挡判断机制,在检测到较大比例遮挡后停止分类器更新,并通过缩放搜索样本的尺寸解决目标尺寸快速变化导致的遮挡检测虚警;对于具有一定运动信息的目标,通过卡尔曼滤波器进行预测解决目标全遮挡后的持续跟踪.实验证明,提出的算法框架对遮挡情形下的目标跟踪具有较高的鲁棒性.

Struck Based Object Tracking under Occlusion

Struck is a vision target tracking algorithm with better comprehensive performance in recent year, but for larger proportion of occlusion or all occlusion, the performance of the algorithm decreases obviously. The analysis of Struck shows that the classifier will learn wrong information when occlusion occurs which result in tracking failure or tracking drifting. Accordingly, an active occlusion detection method is applied to Struck framework in this paper, the classifier update is stopped after large proportion of occlusion is detected and zooming the search samples to reduce the false alarm of occlusion detection caused by size change of the object; as for the object with available motion information, utilize Kalman filter to achieve continuous tracking after total occlusion. The experiment demonstrates that the proposed algorithm framework is of high robustness for object tracking under occlusion.