引入多级匹配机制的多目标连续跟踪方法

Multi-target Tracking Method with Multilevel Matching Mechanism

为提升基于检测的多目标跟踪方法在短暂遮挡时跟踪的准确度,首先提出一种目标多级匹配机制,将检测框根据置信度分为高分框和低分框,使用高分框和之前的跟踪轨迹进行第一级匹配,其次使用低分框和第一次没有匹配上高分框的跟踪轨迹进行第二级匹配,最后使用行人重识别的方法将未匹配到跟踪器和目标检测的结果进行第三级匹配,最终实现短暂遮挡下的连续跟踪。在MOT17数据集上进行验证,结果表明,所提方法 MOTA值达到76.9,可有效降低多目标跟踪ID切换,提升跟踪的连续性。

In this study,a target multi-level matching mechanism is proposed for improving the tracking accuracy of the detection based multi-target tracking method in the case of temporary occlusion.The detection boxes is divided into low score boxes and high score boxes according to the confidence.The high score boxes is used for the first matching with the previous tracking tracks,and the low score boxes is used for the second matching with the tracking tracks not matched at the first time,Finally,the pedestrian recognition method is used to match the track without matching for the third time,and finally realize continuous tracking under temporary occlusion.This method is tested using the MOT17 dataset.Results show that the MOTA of the proposed method is 76.9,which can effectively reduce the ID switching of multi-target tracking and improve the continuity of tracking.