基于运动动态和多层超图关联的多目标跟踪

Multi-target tracking based on motion dynamics and hierarchical hypergraph association

多目标跟踪会面临目标间外观相似、数据丢失、目标间的轨迹交叉以及相机运动等多种问题。文章提出一种基于目标运动动态和多层超图关联的方法,可有效解决以上问题带来的影响。首先,在不使用先验运动模型的前提下,通过利用目标运动动态,获取整个视频内所有目标的运动动态信息,构造轨迹碎片间的相似度比较函数,减小具有相似外观的不同目标之间的误匹配;其次,使用超图关联,对各个轨迹碎片进行全局搜索聚类,使得跟踪问题转化为一个动态搜索超图的超边集问题,优化求解后跟踪系统能够有效处理长时间的遮挡,并且具有较好的鲁棒性。实验表明,在具有挑战性的公共视频序列,该文提出的方法显示了其良好的优越性,能够有效克服目标的复杂运动、相机运动和长时间的遮挡,而这些都是没有任何外观信息的。

Multi-target tracking will encounter many problems,such as appearance similarity among targets,missing data due to targets being out of the field of view or occluded behind other objects,crossing trajectories and camera motion.In this paper,a method based on the target motion dynamics and multi-layer hypergraph association is proposed,which can effectively solve the above-mentioned problems.Firstly,without using the motion model,the motion dynamic information of all the targets in the whole video is obtained by using the target motion dynamics,and the similarity comparison function between the trajectory fragments is constructed to reduce the mismatch between different objects with similar appearance.Then,by using the method of hypergraph association,the global search clustering of each trajectory fragment makes the tracking problem become a hyper edge set problem of a dynamic search hypergraph.The tracker can effectively handle long-term occlusion with more robustness by the optimization process.Experiments show that in the challenging public video sequence,the presented method shows its good superiority in handling complex target motions,non-stationary cameras and long-term occlusions on scenarios where appearance cues are not available.