A Decision-Making Model for Autonomous Vehicles Considering Pedestrian’s Time Pressure Based on Game Theory and Reinforcement Learning

Due to the obvious randomness, pedestrian crossing behavior is hard to predict, which challenges the decision-making of autonomous vehicles (AVs). Recent solutions have been able to adapt to structured road scenes with crossing signals or markings. However, there is still a gap in extending the pedestrian-vehicle interaction (PVI) performance in structured road scenes to unstructured road scenes. Therefore, this paper proposed a vehicle decision-making model considering pedestrian intention based on game theory and reinforcement learning (RL). We designed and conducted a simulation experiment based on a virtual reality platform. Then, leveraging game theory, we established a pedestrian crossing decision-making model considering pedestrian heterogeneity evoked by time pressure (TP). A reward function was developed to enhance driving performance by combining safety, efficiency, and comfort. The RL agent of AVs learns to control the vehicle speed in a pattern that maximizes cumulative rewards through trials and errors by interacting with pedestrians in the simulation environment. The results show that AVs can effectively and safely interact with heterogeneous pedestrians on unstructured roads based on the proposed model. This study contributes to developing AVs that interact better with pedestrians and improve traffic safety, efficiency, and user acceptance of autonomous vehicles.