Intelligent vehicle pedestrian light (IVPL): A deep reinforcement learning approach for traffic signal control

摘要

Deep reinforcement learning (RL) has been widely studied in traffic signal control. Despite the promising results that indicate the superiority of deep RL in terms of the quality of solution and optimality over fixed time signal control, the real-world multi-modal traffic flows, especially pedestrians, are not properly considered nor sufficiently investigated. This study presents a novel deep RL-based adaptive traffic signal model to control the vehicles and pedestrian flows by allocating an equitable green time to each, aiming at minimizing "total user delays" as opposed to "total vehicle delays" dominantly being used in the literature. Our proposed intelligent vehicle pedestrian light (FVPL) method can perform in the absence or presence of pedestrians, especially when there is jaywalking at the intersection, interrupting vehicle flows. To this end, an extended reward function is designed to capture delays due to vehicle-to-vehicle, vehicle-to-pedestrian, and pedestrian-to-pedestrian interactions, as well as red-light delays for vehicles and pedestrians. To evaluate the performance of IVPL, a microsimulation model of an intersection in city of Melbourne is used as a case-study. The real traffic signal parameters of an existing operation system (SCATS) are employed, and the simulation is calibrated using video-based camera data and loop detectors data collected at intersection. The experimental results demonstrate the superiority of the proposed model over fully actuated traffic signal, not only in terms of the quality of optimal solution, but also considering the fact that the proposed model can minimize the "total user delays".