To enhance the safety and efficiency of dynamic obstacle avoidance for unmanned surface vessels (USV), an obstacle avoidance method combining the Velocity Obstacle Method and Deep Q-Network (DQN) is proposed. First, when calculating the traditional velocity obstacle's relative collision region, the future movement information of obstacles is considered. This improvement addresses the issue of obstacle avoidance failure caused by ignoring the real-time position changes of obstacles in the traditional velocity obstacle method. Second, a collision risk coefficient is introduced into the DQN state space, prioritizing obstacles with the highest risk coefficient as avoidance targets, thereby reducing redundancy in state space information. Third, a reward function is redesigned based on the improved velocity obstacle method's obstacle avoidance concept, determining the timing and steering angle for USV obstacle avoidance. This solves the sparse reward problem in traditional DQNs, enhancing their learning efficiency and convergence speed. Finally, to validate the performance of this method, simulation experiments were conducted comparing it with three mainstream obstacle avoidance methods. The results show that this method can provide suitable avoidance directions for USVs, making their navigation paths more economical and safer. Additionally, real-ship experiments confirmed the method's practical engineering value.