Impact of Pedestrian and Vehicle Connectivity on Intersection Performance: A High-Fidelity Simulation with Deep Reinforcement Learning Control

Abstract

Recent studies on the adoption of deep reinforcement learning (DRL) for traffic signal control (TSC) have demonstrated promising outcomes. However, limited research has explored the partial connectivity of diverse agents for DRL-TSC applications for intersections with the presence of pedestrians. Most existing models assume complete detection and communication among agents and often fail to utilize a high-fidelity simulation that includes real heterogeneous data and communication protocols. In this work, we present a holistic co-simulation that integrates high-fidelity Simulation of Urban Mobility (SUMO) and OMNeT++ simulations with real trajectory data of vehicles and pedestrians to evaluate a DRL-based TSC model in a mixed-connectivity setting. The Deep Q-Network model features a novel state representation inspired by the discrete-time state estimation method and is used to explore the effect of partial connectivity of vehicles and pedestrians on DRL models. Real-world trajectory data from the TGSIM dataset are embedded to define route distributions and calibrate behavioral models for vehicles and pedestrians—using the intelligent driver model and social force model, respectively. Experiment results indicate that, by adjusting penetration rates, the DRL model can prioritize the passage of specific agents, improving traffic management efficiency in critical situations. Additionally, the study shows that, even with a partially detected environment, DRL models can outperform traditional traffic control methods within certain penetration ranges. The framework and source code are made publicly available to foster further research into multi-agent communication-aware TSC systems.