Driving behavior modeling is a crucial yet challenging task in the development of traffic simulation systems. Advances in machine learning and data-driven vehicle trajectory extraction technologies have significantly advanced research in this area. However, the performance of such models can be affected by numerous factors often overlooked by the existing methods, including the complexity of real-world road environments and driver characteristics. In this paper, we introduce a novel modeling approach, termed the customized generative adversarial imitation learning (Cus-GAIL) method, designed to capture these complex factors. Our approach incorporates a conditional imitation learning technique that utilizes traffic’s prior knowledge to train a reinforcement learning (RL) model. In addition, we have innovatively developed a collision avoidance mechanism that markedly improves the reliability of microscopic traffic simulation. To address variations in driving styles, we have also created a driver classifier. Moreover, we propose a method for synthesizing small-sample vehicle trajectory data to enhance the RL model’s ability to perceive rare data scenarios. By integrating these components, our model effectively encapsulates a wide range of external and internal factors. To validate the efficacy of the Cus-GAIL method, we employ an unmanned aerial vehicle (UAV) to monitor the two road segments and gather video data of actual vehicle trajectories. The experimental results demonstrate that the Cus-GAIL method outperforms established baselines on both microscopic and macroscopic metrics.
Customized Generative Adversarial Imitation Learning for Driving Behavior Modeling in Traffic Simulation
2025
Aufsatz (Zeitschrift)
Elektronische Ressource
Unbekannt
Metadata by DOAJ is licensed under CC BY-SA 1.0