Physics informed deep generative model for vehicle trajectory reconstruction at arterial intersections in connected vehicle environment

Abstract

Inferring the complete traffic flow time–space diagram using vehicle trajectories provides a holistic perspective of traffic dynamics at intersections to traffic managers. However, obtaining all vehicle trajectories on the road is infeasible. To this end, a novel framework that combines the conditional deep generative model and physics-based car-following model is proposed to reconstruct all vehicle trajectories from sparsely available connected vehicle (CV) trajectories at the intersection. The proposed framework has two novel components: Arrival Generative Adversarial Network (Arrival-GAN) and Trajectory-GAN. The Arrival-GAN reproduces stochastic vehicle arrival patterns by considering the interaction between adjacent intersections (e.g., signal control scheme) and the interaction between multiple vehicles from historical vehicle trajectories, circumventing the conventionally adopted unrealistic assumptions of uniform vehicle arrivals. The Trajectory-GAN model takes the baseline trajectory deduced by the physics-based car-following model as prior information and refines it by dynamically adapting driving behavior in response to the varying traffic conditions in a data-driven manner. This hybrid approach leverages the advantages of data-driven (i.e., flexibility) and theory-driven approaches (i.e., interpretability) complementarily. The proposed framework outperforms conventional benchmark models in the simulated arterial network and the real-world datasets, reconstructing a complete time–space diagram at intersections with markedly enhanced accuracy, particularly in low-traffic-density scenarios. This study showcases the potential of utilizing CV data and physics-informed deep learning to improve our understanding of traffic dynamics, empowering traffic managers with novel insights for efficient intersection management.