Vision-based environment perception and autonomous obstacle avoidance for unmanned underwater vehicle

Abstract

We propose two schemes to solve the problems of visual perception delay and limited obstacle avoidance capability in the autonomous navigation of unmanned underwater vehicles (UUV). In the aspect of visual perception, this paper presents a compressed algorithm based on the UNDERWATER-CUT model for underwater image enhancement and a two-stage compressed YOLOv5s model for object detection, simultaneously reducing the model volume, decreasing the run-time memory footprint, and lowering the number of computing operations, without compromising enhancement effect and recognition accuracy. The 2D coordinates of the obstacles acquired by the visual perception algorithms and the depth values by the depth camera are then merged into real-time navigation information. In terms of obstacle avoidance, this paper proposes the adapted modified guidance vector field (AMGVF) to achieve autonomous navigation. A high-fidelity underwater simulation platform based on AirSim is constructed to jointly validate the above algorithms. Finally, the experimental results on visual perception and obstacle avoidance show that the above methods apply to the UUV well in complex underwater environment.