Intersection Access Decision Making of Robot Guide Dog via a Novel Deformable Attention Transformer-Based Detection Algorithm
Shida Liu, Qingsheng Liu, Li Wang, 蒋红海 Jiang Honghai · 2024
Aiming at the problem of quadruped robotic dogs passing at intersections with low postures and much occlusion, which are not easy to recognize traffic signals, a novel Deformable Attention Transfrmer-based YOLO (DAT-YOLO) detection algorithm is proposed. The proposed DAT-YOLO method enhances the YOLO V5 framework by incorporating a dattention attention mechanism to the backbone section, which empowers the DAT-YOLO algorithm to detect traffic signals with enhanced precision. Moreover, the proposed algorithm was deployed and experimented on the practical Unitree Go1-type quadruped robot. By using the real dataset collected from the practical sidewalk scenario, the effectiveness and applicability of the proposed algorithm is verified via a series of experiments.