It is necessary to plan a feasible path and avoid moving obstacles in real time, while the unmanned surface vehicle (USV) performs tasks in the archipelago environment. This paper focuses on the obstacle avoidance of USV in dynamic environment. A hierarchical framework is proposed for the underactuated dynamic system of USV. Devolution of the dynamic obstacle avoidance from planner to controller is more effective. Rapidly-exploring Random Trees Star is used as a global planner to search for a feasible path, combined with a reactive approach to modulate the control input to avoid moving obstacles in real time. The feasibility of the proposed method is verified by the numerical simulation, which can improve the real-time performance of dynamic obstacle avoidance. Different from the replanning, we adjust the controller to avoid local collision, while adopt the planner to generate a global path.