Research on visual-guided unmanned surface vehicle for water target recognition and tracking

To enhance the capability of Unmanned Surface Vehicles (USVs) in recognizing and locating drowning persons during water rescue operations, this paper presents a target recognition and tracking method that integrates deep learning-based visual features with binocular stereo vision. The system utilizes the YOLOv5 object detection algorithm for rapid localization of drowning persons in images, combined with the Semi-Global Block Matching (SGBM) stereo matching algorithm to generate disparity maps. The three-dimensional position of the target is then calculated based on the triangulation principle. The system is deployed on the Rockchip RK3588 embedded platform, employing software-hardware co-design to optimize the model inference and image processing pipelines. Experimental results demonstrate that the system achieves real-time, high-precision recognition and localization of drowning persons in both laboratory and real-water environments, exhibiting strong robustness and practicality. This work provides reliable visual perception support for water rescue robots.