Research on Vehicle Trajectory Prediction Methods in Urban Main Road Scenarios

In the autonomous driving realm, it is crucial to accurately predict the trajectory of other traffic vehicles around an autonomous vehicle. This paper proposes a novel scheme to predict the trajectory of the target vehicle driving on urban main roads efficiently and accurately. Inspired by the substantial impact of road geometry on vehicle trajectory, we delve into three feature input methods to encode vehicle features. We integrate a multi-head attention mechanism for weighting map features. Following this, we utilize an attention gating mechanism to merge map features with the vehicle’s dynamic features. Spatial and temporal attention mechanisms are employed to extract the spatial-temporal interaction features of the vehicle. We evaluate the proposed method on the Argoverse Motion Prediction dataset and the INTERACTION dataset. Compared to VectorNet, our proposed model achieves a reduction of 10.8% in Average Displacement Error (ADE) and 14.8% in Final Displacement Error (FDE), respectively. The experimental results show that our method achieves better prediction performance than other benchmark methods in various types of urban main road scenarios. The designs of various modules in our model can effectively improve the trajectory prediction model’s performance.