Predicted occupancy map-based online safety verification and resilient motion planning for autonomous driving

• Proposes a method for predicting the future occupancy regions of dynamic traffic participants. • Proposes a resilient safety motion planning method that ensures the autonomous vehicle consistently maintains a safe trajectory. • Enhances the performance of planning strategies by integrating reinforcement learning with imitation learning. • Experimental results indicate that the occupancy region prediction method achieves high accuracy and generalization. Ensuring the safety of Autonomous Vehicles (AVs) continues to present a significant challenge. In order to achieve broader acceptance, it is essential to ensure that autonomous driving systems consistently make safe decisions in all traffic scenarios and are validated for safety compliance. This paper proposes a Predicted Occupancy Map-Based Online Safety Verification and Resilient Motion Planning (POM-SVP) method as a safety layer for current motion planning algorithms. The method aims to verify whether the planned trajectories of autonomous driving systems meet safety requirements and to provide resilient safe trajectories in critical scenarios. By combining explicit and implicit traffic rules, as well as kinematic models of vehicles and pedestrians, the proposed method effectively predicts the future over-approximated occupancy region of surrounding traffic participants to assess whether the intended trajectory of the autonomous vehicle presents any safety conflicts. Furthermore, our method utilizes a Transformer module to provide resilient, conflict-free safety trajectories with respect to the future occupancy of surrounding traffic participants, serving as backup switching solutions to ensure the vehicle always operates within a safe region. Experimental results indicate that the proposed method can significantly reduce the number of traffic accidents while maintaining vehicle motion continuity and intention consistency. By integrating our safety verification algorithm, we significantly outperformed previous research without integration on the widely recognized CARLA leaderboard, with the driving score improving by up to 119.13%, surpassing the current top-ranked entry.