Abstract
<title>Abstract</title> <p>To address the collision avoidance decision-making problem faced by Unmanned Surface Vehicle (USV) during autonomous navigation, this paper proposes an autonomous collision avoidance decision-making method for USV that integrates Unmanned Aerial Vehicle (UAV) visual perception and Deep Reinforcement Learning (DRL).Firstly, the visual image information acquired by UAV is used as the input of the neural network model to realize the perception and representation of the surrounding navigation environment. A Prioritized Experience Replay (PER) mechanism is introduced into the framework of the Dueling Double Deep Q-Network (D3QN) algorithm to optimize the sample utilization efficiency, thereby improving the training efficiency of the algorithm.Secondly, a new type of reward function is designed to guide USVs to reach the target position safely while ensuring that their collision avoidance behaviors comply with the International Regulations for Preventing Collisions at Sea (COLREGs).Finally, to verify the effectiveness and superiority of the proposed method, collision avoidance simulation experiments are conducted under various scenarios. The experimental results show that the proposed method enables USV to make safer and more efficient decision-making behaviors, and exhibits excellent environmental adaptability and potential for engineering applications. Index Terms—USV；Deep reinforcement learning；Collision avoidance</p>