Two-step deep reinforcement learning for traffic signal control to improve pedestrian safety using connected vehicle data
提出一种两步深度强化学习框架，利用联网车辆数据动态调整信号配时，在减少人车冲突和延误的同时平衡行人安全与通行效率，仿真验证有效。
The primary goal of traffic signals control (TSC) is to enhance safety and protect all traffic participants. However, there exists enhancement such as increasing safety for vulnerable road users (VRUs), especially pedestrians. This study proposes a novel two-step traffic signal control framework based on deep reinforcement learning (TSDRL-TSC) to improve pedestrian safety and overall traffic efficiency at intersections. Based on advanced communication technologies of connected vehicles (CV), the TSDRL-TSC acquires the data from real-time traffic conditions and dynamically adjusts traffic signals, aiming to minimize traffic conflicts and delays of pedestrians and vehicles. In the first step, TSDRL-TSC decides whether to use traditional four-signal phases or a modified version considering the protected/prohibited right turn (PPRT) strategy based on pedestrian conditions. In the second step, TSDRL-TSC optimizes the specific control scheme through deep reinforcement learning techniques, selecting the optimal signal phases/actions for the current intersection state to obtain long-term reward returns. The reward function considers the safety and efficiency of all traffic participant, designed to balance the requirement for pedestrian safety, pedestrian efficiency, and vehicle throughput. Simulation experiments at a representative six-lane bidirectional intersection in Changsha City validate the effectiveness of the proposed method. Results demonstrate that (1) TSDRL-TSC significantly reduces pedestrian-vehicle conflicts, jaywalking incidents, and total delays compared to adaptive traffic signal control and PPRT control; (2) TSDRL-TSC presents the potential as a robust solution to enhance pedestrian safety and traffic efficiency for complex urban traffic management.