ASTNAT: an attention-based spatial–temporal non-autoregressive transformer network for vehicle trajectory prediction

Accurate vehicle trajectory prediction is a fundamental prerequisite for downstream tasks like safety analysis and trajectory planning. Despite extensive research efforts for trajectory prediction, challenges still persist in fusing driving intentions and deeply exploring interaction features among vehicles in complex traffic environments. In this paper, an Attention-based Spatial–Temporal Non-Autoregressive Transformer (ASTNAT) network to capture complex features in historical information for vehicle trajectory prediction is proposed. To begin, we present the Historical Trajectory Time Encoder (HTTE) module, designed to capture the dependencies of different timestamps of vehicles. Secondly, interaction features between the target vehicle and surrounding vehicles at timestamp \(t\) $t$ are captured by the Spatial Social Interaction (SSI) module, while the Spatial–Temporal Social Dependence (STSD) module captures interaction features across different timestamps. Thirdly, the Driving Intention Feature Fusion (DIFF) module is used to capture driving intention features. Finally, in the Decoder Output module, future trajectories are generated all at once in a non-autoregressive manner, avoiding the impact of error accumulation in single timestamp iterative prediction output. Experimental results on the NGSIM and HighD datasets indicate that compared to state-of-the-art models, the proposed model exhibits better performance. The Root Mean Square Error (RMSE) of the prediction trajectory at 5 s time horizon is 3.23 m on NGSIM dataset and 1.31 m on the HighD dataset. Furthermore, ablation experiments are conducted to evaluate the performance of each module.