Graph-Guided Trajectory Prediction with Bidirectional Dynamic Interaction Modeling for Autonomous Vehicles

Abstract As a crucial component linking the perception and decision-making modules of automated vehicles, vehicle trajectory prediction plays a key role in improving driving safety and efficiency. In complex driving environments, simply extrapolating future trajectories from historical states is insufficient for accurate prediction, as vehicle interactions can induce significant trajectory variations. In this paper, a graph-guided vehicle trajectory prediction method inspired by bidirectional dynamic interaction is proposed. Given the cyclic coupling of interactions, undirected graphs are constructed based on distance and velocity. Multilayer graph convolutional networks (GCNs) are employed to extract bidirectional interaction features among vehicles at each time step. Building upon this, a temporal correlation extractor based on gated recurrent units (GRUs) is established, extending single-step interactions to continuous-time interactions and capturing dynamic interaction representations. Furthermore, to leverage the improved trajectory prediction accuracy enabled by bidirectional dynamic interactions, relevant features are mapped into a latent space via a conditional variational autoencoder (CVAE), enabling the modeling of continuous and complex driving behaviors. The proposed method achieves an 11.2% improvement in prediction accuracy over the typical method TrajectoFormer on the NISIM dataset. Real-world vehicle experiments conducted on the Dongfeng AX7 and Hongqi EHs3 platforms further validate the practical applicability of the proposed method.