Pedestrian Multi-target Tracking Based on YOLOv3
Sheng Tian, Jun Liu, Yuan-dong Jin, Chao-yu Deng · 2020
Pedestrian multi-object tracking is an important research direction in the fields of computer vision which has been widely applied in intelligent surveillance, human-computer interaction and so on. The traditional method has high computation complexity, slow detection speed and cannot track in real time. To solve this problem, a multi-object tracking method based on YOLOv3 is proposed. It makes use of the accuracy and timeliness of YOLOv3 in detection, and combines with an efficient data association method to track pedestrian targets in real time. More discriminating pedestrian re-identification network is introduced in order to reduce ID switches in tracking, and it alleviates the problem of long term occlusion. The experimental results on datasets show that several indexes have been improved and the proposed method can better adapt to complex scenes and extract more complete target trajectories.