Obstacle Detection and Avoidance Method Using Object Detection Model and Stereo Vision

Abstract Autonomous underwater vehicles (AUVs) navigate and conduct observations underwater without cables. Because AUVs have limited operator involvement during navigation, they must have great autonomy. Therefore, while recognizing their surrounding environment and planning mission routes, AUVs must detect and avoid obstacles. Our proposed method combines an object detection model and stereo vision to recognize the environment. The object detection model is used to discern known and unknown detected objects. Stereo vision can reveal the detected object distance and size. Method for obstacle avoidance is a route generation problem based on planned route and detection information. Dijkstra’s algorithm, a common route search algorithm, and the A* algorithm are usually applied on a two-dimensional plane, but the proposed method extends the A* algorithm to three dimensions for underwater navigation. The proposed method was evaluated by simulation. After preparing a virtual seabed environment using simulator software, the method was evaluated with several placed objects. Findings confirmed that the system can change course to avoid objects recognized as obstacles. The AUV’s on-board computer is expected to be able to process these calculations in real time. Future system evaluation and improvement are planned, along with implementation on an actual AUV and evaluation using actual sea trials.