Integrating Bird’s Eye View Fusion and Reinforcement Learning for Efficient Autonomous Intersection Navigation

The automotive industry has seen rapid advancements with the integration of artificial intelligence (AI), particularly in autonomous driving. While significant progress has been made, navigating complex traffic scenarios, such as intersections, remains challenging due to their dynamic nature. Traditional rule-based systems often struggle to adapt, prompting the need for more advanced solutions. This study aims to enhance autonomous driving at intersections by integrating bird’s eye view (BEV) fusion and reinforcement learning (RL). Using the CARLA simulator, we fine-tune the UNetXST model to fuse multiple camera perspectives into a BEV representation, providing a holistic view of the vehicle’s surroundings. This comprehensive view serves as input for the RL agent, which is trained using the proximal policy optimization (PPO) algorithm to learn optimized driving strategies, avoid collisions, and ensure efficient navigation. Our results show that the proposed framework outperforms baseline techniques.