Research on Collision Avoidance Method of USV Based on UAV Visual Assistance

Collision avoidance technology serves as a critical enabler for autonomous navigation of unmanned surface vehicles (USVs). To address the limitations of incomplete environmental perception and inefficient decision-making for collision avoidance in USVs, this paper proposes an autonomous collision avoidance method based on deep reinforcement learning. To overcome the restricted field of view of USV perception systems, visual assistance from an unmanned aerial vehicle (UAV) is introduced. Perception data acquired by the UAV are utilized to construct a high-dimensional state space that characterizes the distribution and motion trends of obstacles, while a low-dimensional state space is established using the USV’s own state information, together forming a hierarchical state space structure. Furthermore, to enhance navigation efficiency and mitigate the sparse-reward problem, this paper draws on the trajectory evaluation concept of the dynamic window approach (DWA) to design a set of process rewards. These are integrated with COLREGs-compliant rewards, collision penalties, and arrival rewards to construct a multi-dimensional reward function system. To validate the superiority of the proposed method, collision avoidance experiments are conducted across various scenarios. The results demonstrate that the proposed method enables USVs to achieve more efficient autonomous collision avoidance, indicating strong potential for engineering applications.