Social Attention Network Fused Multi-Patch Temporal-Variable Dependency Based Trajectory Prediction for Internet of Vehicles

Vehicle trajectory prediction (VTP) is important for ensuring safe decision-making and planning in Internet of Vehicles (IoV). In complex traffic scenarios, accurate and reliable trajectory prediction requires comprehensive understanding of the interaction behaviors among vehicles. However, existing methods fail to effectively capture vehicle interaction features and fully explore their potential dependencies, limiting improvements in prediction accuracy. To this end, we propose a social attention network fused multi-patch temporal-variable dependency (SAN-FTVD) model to tackle the above problems. In specific, we first design a variable token embedding module (VTEM) to extract the motion state information of vehicles, which independently embeds each variable of vehicle historical data into a variable token. After that, we propose a physical informed vehicle interaction encoder (PI-VIE) to capture vehicle interaction features over continuous time. The encoder is combined with physical priors to encode vehicle interaction features based on the correlations between the variable tokens. Following that, a temporal-variable dependency fusion module (TVDFM) is proposed to extract and fuse the multi-patch temporal and variable dependencies, fully exploring potential dependencies in vehicle interaction features. Numerical results demonstrate that compared with the state-of-the-art model, the proposed model reduces the average prediction root mean square error over 5-second time range by 8% and 7% on two public datasets with 75% less inference cost. Furthermore, extensive ablation experiments validate the effectiveness of the above modules in the model.