ALeader-following Method Based on Binocular Stereo Vision For Quadruped Robots

Abstract

This paper presents a leader-following framework by stereo-vision, which allows quadruped robots to adapt to challenging outdoor environments. Our framework relies on the deeply supervised object detector (DSOD) detection framework and the kernelized correlation filter (KCF) tracker. The algorithm can enhance the cognitive ability of robots and lay a foundation for the autonomy of robots in outdoor environments. Without the use of pre-trained models, DSOD is a framework for training object detectors from scratch, with greater accuracy than state-of-the-art detectors. We propose an improved tracker based on traditional KCF with multi-feature fusion and multi-scale estimation. Pedestrian re-identification and relocation are introduced to mitigate effects of strong vibration, scale variation, occlusion, and illumination variation. To compensate for the possible transient balance loss caused by short tracking drift, the heading angle of the robot is smoothed and filtered. Outdoor experiments with the quadruped robot show the effectiveness and robustness of this framework.