Enhancing Autonomous Driving Safety Model through PRDQN and Zero-Shot segmentation in Real-Time Traffic Scenarios

Safety is a key prerequisite for autonomous driving systems, yet the many unpredictable corner cases on public transportation remain a huge hazard. By definition, a corner case is the presence of unpredictable and relevant objects/categories at the location in question, including sudden traffic accidents, unmarked roadblocks, and so on. To this end, we introduce a perceive everything autonomous approach that can still perceive shapes and categories in real-time traffic scenes with zero-shot learning. In addition, considering the scarcity of corner cases, we implement the DQN algorithm with prioritized experience replay (PER) to effectively balance the empirical equilibrium between corner cases and generic cases. Finally, we designed four different trajectories on CARLA simulator, a real-time simulator for autonomous driving, and compared them with other autonomous driving algorithms to achieve very excellent results. In addition, we perform ablation experimental analyses of our own models to validate the effectiveness of the segmentation everything algorithm module and the DQN module with prioritized experience replay.