Real-Time Obstacle Detection and Localization for USV Using Stereo Vision and Edge Computing

Unmanned Surface Vessel (USV) is widely used in various fields, and obstacle avoidance is a key technology for USV. Accurately detecting the position and type of obstacles is crucial for obstacle avoidance. Active sensors can accurately detect the position of obstacles but cannot distinguish their types, necessitating the use of vision sensors. Therefore, this paper employs stereo vision and proposes a stereo obstacle detection and localization method, based on SGBM and YOLOv5. The SGBM algorithm generates depth information, while the YOLOv5 algorithm detects the position of obstacles in the image. To address the issue of errors in measuring the continuous position of obstacles, this paper uses a Kalman filter to smooth the trajectory. Experimental results show that the proposed method improves obstacle localization accuracy by approximately 38% compared to raw measurements. Besides, considering the energy constraints and computing capabilities of lightweight USV, which cannot meet the demands of processing vision tasks, this paper introduces an edge-device collaboration architecture, by offloading the USV computing tasks to edge nodes, thereby extending the computing capacity and energy efficiency of the USV. Experimental results indicate that the proposed architecture meets the real-time obstacle avoidance requirements of USV.