This paper addresses the issues of insufficient obstacle avoidance efficiency in traditional path planning algorithms (such as A* and RRT) for drones in complex environments, as well as the limited local obstacle avoidance flexibility of single deep reinforcement learning methods (such as PPO). It proposes a hybrid algorithm that integrates PPO with the Dynamic Window Algorithm (DWA). The main contributions are: (1) A configurable and randomized 3D training environment is constructed using Airsim and Unreal Engine, supporting the parametric generation of dynamic and static obstacles, which enhances training stability. (2) PPO is used for global path planning, while DWA generates local speed commands and provides feedback on feasible areas, forming a local guidance mechanism to enhance real-time obstacle avoidance capabilities. (3) A curriculum learning strategy is employed to optimize the reward function and guidance range in stages, accelerating model convergence and improving robustness in complex environments. This method significantly improves the autonomous navigation and obstacle avoidance performance of drones in dynamic scenarios.