Aiming at the problems faced by unmanned vehicles in dynamic environment, such as the uncertainty of obstacle movement, the limited time delay of online re planning, and the difficulty of taking into account the safety and smoothness of trajectory, this paper proposes a real-time path planning method based on improved deep reinforcement learning. Based on the local dynamic occupation representation, the method integrates risk information such as relative distance, relative speed, collision time and obstacle density, and introduces attention weighting to enhance the perception of key targets; At the decision-making level, improved SAC and priority experience playback are used to improve the stability of continuous motion learning; In the execution layer, the safety action correction and MPC smooth optimization are combined to improve the trajectory executability. Four kinds of dynamic scenes were constructed based on carla-ros joint environment, and compared with A*, DWA, PPO and original SAC. The results show that the proposed method has better comprehensive performance in terms of success rate, collision rate, path length, minimum safe distance and online planning delay, while maintaining good robustness and generalization ability in the absence of speed, higher obstacle density and noisy conditions.
