High-fidelity simulation of pedestrian-vehicle dynamic interaction in complex urban traffic scenarios is identified as a critical unsolved technical challenge in the autonomous driving field. To address this challenge, a multi-agent game information adversarial inverse reinforcement learning (MA-GI-AIRL) model was proposed in this work. First, the logit stochastic best-response equilibrium was adopted to model suboptimal decision-making behaviors under bounded rationality, and a structured game utility function was constructed by using kinematic indicators such as time-to-collision (TTC), post-encroachment time (PET), velocity, and acceleration to characterize interaction safety, traffic efficiency, and motion comfort, respectively. Second, a mutual information maximization mechanism was introduced to decouple heterogeneous behavioral preferences. Finally, real-world traffic data collected by unmanned aerial vehicles was utilized for model training. An interpretable reward function and corresponding behavioral policies were learned and recovered. The research results show that the proposed MA-GI-AIRL model outperforms the baseline models in terms of both trajectory fitting accuracy and the rationality of kinematic indicators. The average displacement error (ADE) of the model is 0.757 m for pedestrian trajectories. A high-fidelity pedestrian-vehicle interaction simulation environment is provided for development and evaluation of autonomous driving systems by the validated model.