An Enhanced Driving Trajectory Prediction Method Based on Generative Adversarial Imitation Learning

Trajectory prediction stands as a fundamental technology in the development of seamless vehicle-infrastructure collaboration systems, tasked with anticipating the immediate and extended path trajectories of all vehicles on the road to enable safer and more accurate driving decisions. In pursuit of boosting the accuracy of these predictions, this paper utilizes a graph-based structural methodology to construct a highly-detailed rendering of the driving scenario, embedding an agent-centric modeling technique to formulate a probabilistic motion model for vehicles. Furthermore, it embraces Generative Adversarial Imitation Learning (GAIL) to ingeniously craft driving tactics, ultimately generating a spectrum of multi-modal predicted trajectory options. Simulations conducted on the nuScenes motion prediction dataset demonstrate that the proposed method generates trajectories that align closely with the inherent traits of actual road scenarios, exhibiting superior accuracy compared to extant methods. These results underscore the promise of the technique in enhancing the reliability and predictive capability of trajectory forecasts in complex traffic environments.