Game-Generative Adversarial Imitation Learning for Pedestrian Simulation During Pedestrian-Vehicle Interaction

In this paper, we propose a game-generative adversarial imitation learning (G-GAIL) method to simulate realistic pedestrian-vehicle interaction, which takes advantage of both game theory (GT) and generative adversarial imitation learning (GAIL). Pedestrians are modeled as agents. The interaction between pedestrians and vehicles is described as a Markov decision process (MDP). By analyzing real pedestrian-vehicle interaction trajectory data, the game mechanism of pedestrian-vehicle interaction is inferred and the pedestrian's crossing decision is analyzed. The game utilities of pedestrians and vehicles are integrated into the decision generation of GAIL to obtain a pedestrian simulation model that is sensitive to the pedestrian-vehicle game. We compare both simulated and real values of pedestrian behavioral decisions and pedestrian-vehicle interaction parameters. The results show that the developed model can accurately simulate pedestrian behavior under different decision-making conditions and provide more realistic pedestrian-vehicle interaction scenarios for the simulation testing of autonomous driving.