基于YOLOX和重识别的行人多目标跟踪方法

Pedestrian Multi-target Tracking Method Based on YOLOX and Person Re-identification

针对在拥堵场景下多目标跟踪身份频繁切换的问题,该文提出了一种融合行人重识别任务与目标检测任务的联合网络。在YOLOX检测算法上添加重识别(Re-identification)分支,获得含有重识别特征的行人检测框;在ByteTrack跟踪算法的检测框与预测框特征匹配的基础上,利用重识别特征弥补ByteTrack网络在匹配过程中行人外观特征缺失的问题,并结合行人运动特征,进一步提升特征匹配的准确率,减少身份切换次数。在公开数据集MOT17上进行实验,改进后的网络m AP提升2.6%,达到了95.4%,不同尺寸的mAP与mAR均获得明显提升,运行效率几乎保持不变。

To solve the problem of frequent pedestrian occlusion in realistic congestion scenarios by multi-target tracking algorithm. In this paper we proposed a joint network which integrates the two tasks of target detection and pedestrian re-identification. We added a re-identification model to YOLOX to obtain pedestrian detection box with reidentification features. On the basis of the feature matching of ByteTrack between the detection frame and the prediction frame,pedestrian re-identification is used to make up for the lack of pedestrian appearance features in the matching process of the ByteTrack,with combination of pedestrian motion features to improve the accuracy of feature matching and reduce the frequency of identity switching further. In experiment section we compared our enhanced model with ByteTrack on MOT17 which increases the tracking accuracy by 2.6% to 95.4%,the tracking accuracy and recall rate between different sizes are significantly improved,while efficiency remains almost unchanged.