Real-Time Pedestrian Tracking System Based on Depth Camera for Quadruped Robot

Quadruped robots have a wide range of applications now because of their strong adaptability to unstructured terrain. In the field of service robots, quadruped robots can track pedestrian users to provide more personalized services. This paper proposes a real-time depth camera-based pedestrian detection and tracking system for a quadruped robot. Firstly, YOLOv5 method is used for target detection, and then the DeepSORT method is used to collect real-time pedestrian features to complete pedestrian tracking. In order to reduce the adverse effects of target occlusion and fast movement on real-time tracking, this paper improves the performance of pedestrian feature extraction network by adding the CBAM attention module and verifies its superiority on the Market-1501 dataset. Compared with the original network, the Rank-1 accuracy increases by 6.3%, and the mean Average Precision (mAP) increases by 11.9%. Finally, a controller is designed based on kinematic method, and its stability is guaranteed by Lyapunov method.