Intelligent vehicle pedestrian tracking based on YOLOv3 and DASiamRPN

In the field of robot and autopilot, aiming at the difficulty that the traditional single target pedestrian tracking algorithm needs manual frame selection, this paper puts forward a pedestrian tracking algorithm which blends YOLOv3 and DASiamRPN to realize the designated pedestrian tracking. The algorithm automatically selects the tracking target through YOLOv3, and then uses the tracking target as the template to input into DASiamRPN for tracking. The real-time pedestrian tracking experiment based on the Kinect camera was carried out, and the tracking effects of SiamRPN, Deepsort, SiamFC and Mosse+KCF algorithms were compared. The experimental results show that the optimal running speed of DASiamRPN is 34fps while ensuring the tracking accuracy. Finally, the algorithm has been applied to Bulldog robot based on ROS, and the pedestrian tracking is realized successfully, which verifies the real-time and robustness of the algorithm.