Scholarlink

<p>Aiming at the technical bottlenecks of strong dependence on high-precision maps and high complexity of long-term time-domain modeling in autonomous driving, this paper proposes a trajectory prediction model (DynSTGF-TP) based on spatio-temporal coding and dual-channel dynamic graph fusion. This model adopts a two-stage heterogeneous architecture: 1) The short-term motion features and cross-time step global correlations are synchronously extracted through the LSTM-Transformer hybrid encoder (LSTrans); 2) Design a dual-channel dynamic graph fusion Module (DC-DGF), and utilize learnable weights to dynamically balance local geometric constraints (EE-GCN) and the global interaction mode (TransformerConv). The evaluation results on the Argoverse dataset show that this model still achieves superior performance without high-precision map input: In the single-trajectory prediction task, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(minAD{E_{K = 1}}\)</EquationSource> </InlineEquation> reaches 1.80, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(minFD{E_{K = 1}}\)</EquationSource> </InlineEquation> is 4.07, and <i>MR</i> is 0.63, outperforming advanced models such as CRAT; In the multimodal prediction task, <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(minAD{E_{K = 6}}\)</EquationSource> </InlineEquation> is 1.06, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(minAD{E_{K = 6}}\)</EquationSource> </InlineEquation> is 1.82, and <i>MR</i> Is reduced to 0.26. Even compared with advanced models with maps, it is not inferior at all, demonstrating excellent trajectory diversity capture ability.</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

Trajectory Prediction via Spatio-Temporal Encoding and Dynamic Graph Fusion

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Hai-Sheng Li,
  • Hui-Jia Ma,
  • Haizhen Liu,
  • Shuxiang Song,
  • Cong Hu,
  • Xiaohu Tang

摘要

Aiming at the technical bottlenecks of strong dependence on high-precision maps and high complexity of long-term time-domain modeling in autonomous driving, this paper proposes a trajectory prediction model (DynSTGF-TP) based on spatio-temporal coding and dual-channel dynamic graph fusion. This model adopts a two-stage heterogeneous architecture: 1) The short-term motion features and cross-time step global correlations are synchronously extracted through the LSTM-Transformer hybrid encoder (LSTrans); 2) Design a dual-channel dynamic graph fusion Module (DC-DGF), and utilize learnable weights to dynamically balance local geometric constraints (EE-GCN) and the global interaction mode (TransformerConv). The evaluation results on the Argoverse dataset show that this model still achieves superior performance without high-precision map input: In the single-trajectory prediction task, \(minAD{E_{K = 1}}\) reaches 1.80, \(minFD{E_{K = 1}}\) is 4.07, and MR is 0.63, outperforming advanced models such as CRAT; In the multimodal prediction task, \(minAD{E_{K = 6}}\) is 1.06, \(minAD{E_{K = 6}}\) is 1.82, and MR Is reduced to 0.26. Even compared with advanced models with maps, it is not inferior at all, demonstrating excellent trajectory diversity capture ability.