Urban Vehicle Trajectory Generation Based on Generative Adversarial Imitation Learning

With the rapid development of smart cities, the collection of vehicle trajectory data through sensors has increased significantly. While many studies have utilized calibrated physical car-following models (CFM) and machine learning techniques for trajectory prediction, these approaches often falter in complex, dynamic traffic scenarios. Addressing this gap, this paper introduces <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PS-TrajGAIL</i>, a generative adversarial imitation learning framework tailored for urban vehicle trajectory generation. Contrary to conventional discriminative models, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PS-TrajGAIL</i> employs a generative model to capture the inherent distribution of urban vehicle trajectories. This framework models the tasks of trajectory generation as a partially observable Markov decision process based on imitation learning. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PS-TrajGAIL</i>’s architecture features a generator, which simulates vehicle behavior to produce synthetic trajectories, and a discriminator that distinguishes between authentic and generated trajectories. In addition, the driving policy within the generator is fine-tuned using the Trust Region Policy Optimization (TRPO) algorithm, ensuring safety in vehicle driving. Experimental evaluations on both synthetic and real-world datasets highlight that <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">PS-TrajGAIL</i> notably surpasses existing baselines and state-of-the-art approaches in trajectory generation.