A Framework for Automatic Following Based on Visual Tracking for Quadruped Robots

This paper proposes a TOF camera-based auto-matic following framework for quadruped robots, enabling them to track multiple categories of targets and avoid obstacles. Our framework consists of a single target tracking module, a coordinate transformation module, and a navigation module. Specifically, Mixformer is used to obtain the pixel coordinates of the target in the image, and the principles of pinhole imaging are employed to convert these coordinates, along with the depth image, into 3D coordinates and an environmental point cloud. Based on the target's 3D coordinates and the environmental point cloud, the navigation module plans a real-time tracking path that avoids obstacles. Our framework can achieve real-time and stable tracking and obstacle avoidance without additional tags. Experimental validation has confirmed the effectiveness of our framework.