Autonomous obstacle avoidance for UAVs in dynamic obstacle-dominated complex environments must address time-varying local collision risks from multiple directions under the constraints imposed by local sensing, environmental uncertainty, execution safety, and limited onboard computation. Planning-based methods often require frequent replanning or explicit obstacle prediction, whereas conventional reinforcement learning policies may produce myopic decisions under partial observability. To address these limitations, this study proposes a dynamic obstacle-avoidance framework that combines a temporal LiDAR representation with safety-aware action correction in recurrent reinforcement learning. Multi-layer LiDAR observations are constructed using sector-wise minimum pooling. Adjacent two-frame stacking and a CNN-LSTM architecture are then used to extract local geometric structures and short-term dynamic cues, and a velocity-control policy is optimized using Recurrent PPO. In addition, a smooth velocity-projection safety shield is introduced to modify policy outputs and reduce collision risk during both training and policy execution. Experiments conducted in mixed static–dynamic obstacle scenarios based on Gym-PyBullet-Drones show that the proposed method achieves an average success rate of 91.9% across four test configurations, with an average online computation time of 0.78 ms. Ablation studies further support the contributions of two-frame observations, LSTM-based temporal modeling, and the safety shield.