【目的】针对监控视频跨域环境下的多目标跟踪画面中,频繁遮挡、轨迹开始终止时刻未知、目标太小、目标间交互、表观相似以及摄像头视角变化等问题,提出一种改进的多目标跟踪算法。【方法】该算法最大化利用低分检测对象,将未匹配的低分对象进行二次匹配,目标跨域后,依据摄像头拓扑排序规则,以及相邻摄像头的未匹配跟踪轨迹,同时对检测器YOLOv5算法进行优化改进,通过信息流的层层递进,有效解决多尺度问题和小目标信息提取不充分等问题,在相邻的摄像头中快速匹配到跟踪对象,以提高跨域环境下特定多目标跟踪的精度。【结果】对比消融试验表明,本改进算法MOTA达到了62.8%,IDswitch也显著降低。

【Purposes】Multi-target tracking in the cross-domain environment of surveillance  video is a very important and challenging task in intelligent security. The difficulties of this task lie in  the frequent occlusion between the objects in video frame, the unknown start and end time of the tra‐ jectory, the too small sized targets , the interactions between the objects, the apparent similarity, and  the camera angle changes. In view of the frequent occlusions and apparent similar problems, an im‐ proved multiple target tracking algorithm is put forward. 【Methods】 With the maximum use of low  detection object, secondary matching is performed on the unmatched low objects. For target cross- domain, camera’s topological sort rules, adjacent cameras un tracking trajectory, as well as the detec‐ tor YOLOv5 algorithm improvement and the layer-to-layer transfer of information streams, effec‐ tively address the multi-scale problems and the unsufficient information extraction problems for small  objects, promptly match the tracking objects in adjacent camera, thus improving . the accuracy of  multi-target tracking in cross-domain environment.【Findings】 In the comparative ablation tests, the  MOTA value of the improved algorithm reached 62.8%, and the IDswitch value was also signifi‐ cantly reduced.