Dynamic Traffic Signal Control for Pedestrian Safety: A Bayesian EVT and Deep Reinforcement Learning Approach

Pedestrian safety remains a major concern in urban transportation systems. Signal timing strategies such as Leading Pedestrian Intervals (LPIs) have proven to be effective but are typically static and do not account for dynamic traffic conditions. This study develops a safety-based Dynamic Traffic Signal Control (DTSC) framework that integrates pedestrian crash risk into the control process. The proposed framework combines a Bayesian hierarchical Extreme Value Theory (EVT) Peak-over-Threshold (POT) model with a Double Deep Q-Network (DQN) algorithm to dynamically implement LPIs in a signalized intersection in Vancouver, Canada. Results show that the proposed safety-based DTSC reduces both crash risk and conflict frequency compared to conventional actuated control; however, these improvements are accompanied by moderate increases in vehicle delay. Overall, the findings demonstrate that EVT-based crash risk can be effectively used for real-time signal control, enabling dynamic LPIs that improve pedestrian safety while maintaining acceptable mobility levels.