Abstract
With the increasing integration of Connected and Automated Vehicles (CAVs) and Human-Driven Vehicles (HDVs) in urban traffic systems, along with highly variable pedestrian crossing demands, traffic management faces unprecedented challenges. This study introduces an improved adaptive signal control approach using an enhanced dual-layer deep Q-network (EXP-DDQN), specifically tailored for intelligent connected environments. The proposed model incorporates a comprehensive state representation that integrates CAV-HDV car-following dynamics and pedestrian flow variability. Additionally, it features an improved MC Greedy exploration strategy and prioritized experience replay, enabling efficient learning and adaptability in highly dynamic traffic scenarios. These advancements allow the system to dynamically adjust green light durations, phase switches, and pedestrian phase activations, achieving a fine balance between efficiency, safety, and signal stability. Experimental evaluations underscore the model's distinct advantages, including a 26.9% reduction in vehicle-pedestrian conflicts, a 31.83% decrease in queue lengths, a 32.52% reduction in delays compared to fixed-time strategies, and a 35.17% reduction in pedestrian crossing wait times. Furthermore, EXP-DDQN demonstrates significant improvements over traditional DQN and DDQN methods across these metrics. These results underscore the method's distinct capability to address the complexities of mixed traffic scenarios, offering valuable insights for future urban traffic management systems.