A Decision-Making Approach for Complex Unsignalized Intersection by Deep Reinforcement Learning

Decision-making for automatic vehicles at unsignalized intersections with dense traffic is one of the most challenging tasks. Due to the complex structure and frequent traffic accidents, traditional rule-based methods struggle to address this issue flexibly and often produce suboptimal strategies. Recently, deep reinforcement learning (DRL) has garnered significant attention for its exceptional performance in decision-making problems. We propose a local attention safety deep reinforcement learning (LA-SRL) decision-making method for ego vehicle right-turns at unsignalized intersections. LA-SRL enables paying attention to different states of social vehicles within complex traffic environments and effectively deals with the impact of surrounding vehicles on ego vehicle. This contributes to enhancement of safe driving efficiency. To further balance the safety and efficiency of decision-making for ego vehicle at unsignalized intersections with dense traffic flow, we design a safety-reward function composed of risk reward and avail reward. The safety-reward function enables ego vehicle to promptly navigate out of high-risk areas, meanwhile avoiding collisions and reducing waiting periods. Finally, we evaluate our method in the CARLA simulator. The results demonstrate that LA-SRL outperforms state-of-the-art methods, achieving a remarkable success rate of 98.25% and reducing the average time to 6.6 seconds.