Proactive Urban Traffic Safety: Integrating Bayesian Risk Assessment and Reinforcement Learning for Intelligent Urban Infrastructure

Urban traffic faces challenges in balancing congestion mitigation and accident prevention, particularly under adverse environmental conditions. Traditional traffic signal control methods primarily focus on optimizing vehicle flow but often lack proactive risk assessment mechanisms to prevent collisions before they occur. This study proposes a Bayesian network (BN) and deep Q-network (DQN) integrated framework for dynamic risk assessment and adaptive traffic signal control, aiming to enhance both safety and efficiency at intersections. The BN models real-time environmental and traffic conditions to estimate accident risk probabilities, while the DQN-based reinforcement learning optimizes signal phase adjustments to minimize both congestion and collision likelihood. A simulation-based evaluation was conducted, testing the framework’s effectiveness. Sensitivity analysis was performed to assess the impact of key parameters on accident risk and traffic flow. Results demonstrate that the proposed BN + DQN framework reduces collision risk while maintaining a shorter queue length compared to fixed-time and rule-based adaptive controls. Additionally, safety-prioritized signal timing adjustments result in minor trade-offs in travel delay but significantly lower accident probabilities. These findings highlight the feasibility of integrating real-time risk assessment into adaptive signal control, paving the way for safer urban intersections.