关键词: 车辆工程；轨迹预测;PNA;SE-MoE;时空交互;长时预测

Abstract: Due to the lack of degree-based global scaling, the current trajectory prediction models partially loss the effectiveness of spatial feature representation, and the direct concatenation and decoding of spatial-temporal features results in insufficient attention to primary features. Therefore, a vehicle trajectory prediction model based on PNA spatial encoding and SE-MoE fusion of spatiotemporal features was proposed. Temporal features were extracted by a residual-based gated recurrent unit, then spatial features were extracted by spatial encoder, and finally an improved hybrid expert strategy was adopted to fuse and decode spatiotemporal features, realizing the vehicle trajectory prediction. Specifically, based on the extraction of vehicle trajectory spatial features based on PNAConv and GATv2Conv, the degree-based global scaling was employed to enhance the consistency of spatial feature representation, while the attention of key information was enhanced by dynamic attention mechanism. Based on SE-MoE for spatiotemporal interaction feature fusion, the input feature space was divided into three regions with different expression patterns, and the main feature extraction was carried out jointly through routing and shared expert networks. In contrast to traditional concatenation decoding, the SE-MoE-based fusion enhances the expression of trajectory motion trends. Tests on the NGSIM dataset show that compared with models such as Transformer-GAT, S-LSTM and ST-LSTM, the proposed model performs the best in long-term prediction tasks, reducing 5-second root mean square errors by 23.8%, 62.8% and 56.5% respectively, and performs well in short-term prediction tasks. Tests on the Bahe West section of the Xi’an Ring Expressway demonstrate that the proposed model has good generalization capability in real-world scenarios. The 1-second and 5-second root mean square errors for straight-driving samples are 0.78 m and 6.92 m, respectively, while those for lane-changing samples are 0.85 m and 3.29 m, respectively.

Key words: vehicle engineering; trajectory prediction; PNA; SE-MoE; spatial-temporal interaction; long-term prediction