一种基于嵌入空间的防遮挡的多目标跟踪算法

在视频监控中准确识别发生遮挡的目标是视频监控系统中面临一个难点,现在流行的运动目标跟踪算法由于基于检测之上,无法准确判断相互重叠之间的关系。提出一种基于基于嵌入空间识别遮挡关系的方法,在此基础上实现多目标的跟踪。具体的方法是构建遮挡关系函数明确表示发生遮挡目标之间的关系,并将之合并到跟踪算法框架内。同时,对联合状态预测函数进行分步处理,使得多目标的跟踪问题变成在嵌入空间中的分类问题。实验结果表明算法能在有遮挡情景下有效实施目标跟踪,并且,定量分析表明与其他算法相比提高一个数量级。

Multi object tracking algorithm under anti occlusion based on embedding space

It is very difficult to track multi objects while interacting with each other in video surveillance .In this paper a multi object tracking under occlusion based on graph embedding subspace method is presented .In this method , the relationship function between multiple objects is explicitly defined and integrated into the whole track-ing framework.The multi-object separating is viewed as one versus rest classification problem through a new decen-tralized strategy.The experimental results show that the performance of multi object tracking under occlusion is en-hanced effectively.