Multi-target Pedestrian Tracking Based on YOLOv5 and DeepSORT
Yuhan Wang, Han Yang
Abstract
Target detection and tracking are crucial in autonomous driving and intelligent transportation, especially in pedestrians and traffic safety. The dynamic pedestrian tracking algorithm using YOLOv5 and DeepSORT is proposed to improve accuracy and robustness, based on the classical tracking-by-detection approach, to achieve real-time monitoring and tracking of pedestrians in the video. The YOLOv5 algorithm is used for pedestrian detection in the video. The YOLOv5 algorithm is used to detect the pedestrian targets in the video and calculate the pedestrian distance to assess the risk. The detected pedestrian features are tracked in real-time using the DeepSORT algorithm to complete the pedestrian trajectory tracking. The experimental results show that the method is effective in detecting pedestrians when coping with the problems of abrupt and rapid movement, occlusion, and appearance deformations, which is suitable for online real-time tracking of pedestrians.