关键词: 交通工程；轨迹预测；图注意力网络；时空特征融合；自动驾驶

Abstract: Vehicle trajectory prediction is one of the key technologies in intelligent transportation systems and autonomous driving. To address the issues of insufficient attention allocation precision and inadequate spatiotemporal feature fusion in existing vehicle trajectory prediction models under complex vehicle interaction scenarios, a multimodal trajectory prediction model incorporating dynamic perception masking and cross-attention was proposed. Firstly, a spatial interaction feature extraction module based on graph attention network (GAT) was constructed, and a novel dynamic perception masking mechanism was introduced, which dynamically adjusted attention weights according to the vehicle’s own speed and the relative distance to surrounding objects, thereby enabling adaptive focusing on critical interactive objects. Secondly, to facilitate deep coupling of spatiotemporal features, a spatiotemporal feature fusion module based on bidirectional cross-attention was designed, which effectively captured the deep nonlinear dependencies between the two through bidirectional querying and enhancement of temporal and spatial feature streams. Finally, an intention-aware trajectory decoder utilized a gated recurrent unit (GRU) to generate the parameters of bivariate Gaussian distribution describing the trajectories, thereby achieving multimodal trajectory prediction. Experimental results on the public NGSIM and HighD datasets demonstrate that, compared to baseline models, the proposed model exhibits superior performance across all evaluation metrics. The root mean square error (ERMS) in the 5s prediction time domain is reduced by 11.7% and 15.8% on the NGSIM and HighD datasets, respectively, confirming the effectiveness and robustness of the proposed model.

Key words: traffic engineering; trajectory prediction; graph attention network; spatio-temporal feature fusion; autonomous driving