Abstract:

To address the shortcomings of existing motorway vehicle trajectory prediction methods which overlook differences in driving styles and the future interactions between trajectories of different modes. This paper proposes a cross-modal trajectory prediction model based on a dual-hypergraph Transformer for driving styles. The method first quantifies driving styles by extracting vehicle kinematic features and calculates a similarity matrix between vehicles; it then constructs a dual-hypergraph Transformer, using a homogeneous hypergraph to model the group interactions of vehicles with similar driving styles, and a heterogeneous hypergraph to capture the interaction relationships between spatially neighbouring but stylistically dissimilar vehicles; Finally, a style-aware cross-modal interaction module is designed, utilising a multi-head attention mechanism to dynamically adjust the response weights of different driving styles to the interaction scenario, thereby enabling multi-modal trajectory generation. Experiments on the HighD and ExiD datasets demonstrate that this method achieves RMSE values of 1.69 m and 1.66 m, respectively, over a 5-second prediction horizon. Compared with state-of-the-art graph neural network models such as IPF-GCN, it achieves superior performance in terms of average RMSE. On the ExiD dataset, the average prediction accuracy improved by 30.6% and 12.8% compared to the MMnTP and POVL models, respectively. The experimental results demonstrate that the proposed model can effectively capture the interaction relationships within a group of vehicles and significantly enhance the accuracy and stability of vehicle trajectory prediction.

Key words: intelligent transportation, trajectory prediction, driving style, hypergraph, Transformer, cross-modal interaction