Visual Risk-Aware Decision-Making Method for Autonomous Driving

As autonomous driving technology advances, the ability to accurately perceive risks and execute efficient behavioral strategies in real-world traffic scenarios has become critical for performance evaluation. Decision-making algorithms are central to ensuring safety and effectiveness. Addressing the shortcomings of existing systems in comprehensively accounting for risk factors in complex traffic environments, this paper proposes a visual risk perception-based autonomous driving decision model. By integrating visual risk quantification with deep reinforcement learning, the framework evaluates diverse environmental risks to make safer, more rational decisions across varying risk conditions. First, BEVD4D object detection model processes image inputs to achieve vision-based speed estimation. Second, to overcome traditional risk assessment limitations in capturing traffic complexity, a driving risk field model is introduced, considering lane lines, static and dynamic obstacles, and employing a "pseudo-distance"metric to better represent real-world complexity. This model underpins the design of subsequent decision strategies. Finally, the risk field model is embedded into a deep reinforcement learning framework by expanding the state space and reward function of the Soft Actor-Critic algorithm, enhancing risk-aware decision-making. Experiments on the CARLA simulation platform demonstrate the effectiveness and robustness of the method in urban streets and complex intersections, resulting in safer, more adaptive, and interpretable autonomous driving decisions.