Scholarlink

<p>Large-scale cyber-physical transportation systems require efficient multi-vehicle trajectory prediction for real-time traffic management, traffic digital twins, and simulation-based safety assessment. This paper proposes a spatiotemporal interactive fusion dual-graph convolutional network, termed STIF-DGCN, for scene-level highway vehicle trajectory prediction. The model predicts the future trajectories of all observed vehicles in a scene within a single forward pass, supporting parallel inference and large-scale traffic analytics. STIF-DGCN constructs temporal and spatial graphs to capture motion evolution and inter-vehicle interactions. The spatial graph is constructed as a dense candidate graph with a local interaction mask, which reduces redundant interactions from weakly related distant vehicles. Normalized distance- and displacement-derived kernel priors are further fused with attention-derived interaction representations to improve physically meaningful graph construction. Experiments on NGSIM and highD show that STIF-DGCN achieves competitive prediction accuracy over multiple horizons against representative baselines. STIF-DGCN obtains an average RMSE of 1.35&#xa0;m and a 5&#xa0;s RMSE of 2.48&#xa0;m on NGSIM, and an average RMSE of 0.24&#xa0;m and a 5&#xa0;s RMSE of 0.40&#xa0;m on highD. Additional analyses, including fusion operator comparison, perception-range sensitivity, density-stratified evaluation, component ablation, and maneuver-specific testing, further support the effectiveness of the proposed interaction modeling strategy. These results suggest that STIF-DGCN can serve as an efficient and interpretable prediction module for large-scale highway traffic forecasting and real-time decision support. Future extensions will focus on improving its adaptability to more complex and multimodal traffic scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

STIF-DGCN: a spatiotemporal interactive fusion dual-graph convolutional network for multi-vehicle trajectory prediction

  • 期刊论文

  • 发表日期：

    2026年8月2日

查看全文
  • Lujiao Li,
  • Jiafu Wang,
  • Long Chen,
  • Yongbin Hu,
  • Zifeng Liu

摘要

Large-scale cyber-physical transportation systems require efficient multi-vehicle trajectory prediction for real-time traffic management, traffic digital twins, and simulation-based safety assessment. This paper proposes a spatiotemporal interactive fusion dual-graph convolutional network, termed STIF-DGCN, for scene-level highway vehicle trajectory prediction. The model predicts the future trajectories of all observed vehicles in a scene within a single forward pass, supporting parallel inference and large-scale traffic analytics. STIF-DGCN constructs temporal and spatial graphs to capture motion evolution and inter-vehicle interactions. The spatial graph is constructed as a dense candidate graph with a local interaction mask, which reduces redundant interactions from weakly related distant vehicles. Normalized distance- and displacement-derived kernel priors are further fused with attention-derived interaction representations to improve physically meaningful graph construction. Experiments on NGSIM and highD show that STIF-DGCN achieves competitive prediction accuracy over multiple horizons against representative baselines. STIF-DGCN obtains an average RMSE of 1.35 m and a 5 s RMSE of 2.48 m on NGSIM, and an average RMSE of 0.24 m and a 5 s RMSE of 0.40 m on highD. Additional analyses, including fusion operator comparison, perception-range sensitivity, density-stratified evaluation, component ablation, and maneuver-specific testing, further support the effectiveness of the proposed interaction modeling strategy. These results suggest that STIF-DGCN can serve as an efficient and interpretable prediction module for large-scale highway traffic forecasting and real-time decision support. Future extensions will focus on improving its adaptability to more complex and multimodal traffic scenarios.