交通运输系统工程与信息

交通运输系统工程与信息 ›› 2025, Vol. 25 ›› Issue (4): 126-136.DOI: 10.16097/j.cnki.1009-6744.2025.04.013

• 智能交通系统与信息技术 • 上一篇    下一篇

基于异构多图时空融合的长时域车辆轨迹预测

陈峥，张景，陈博闻，李春宇，郭凤香，魏福星*   

  1. 昆明理工大学，交通工程学院，昆明650500
  • 收稿日期:2025-05-06 修回日期:2025-06-22 接受日期:2025-07-07 出版日期:2025-08-25 发布日期:2025-08-25
  • 作者简介:陈峥(1982—)，男，山东潍坊人，教授，博士。
  • 基金资助:
    国家自然科学基金(52272395)；云南省基础研究项目(202401AS070118)。

Heterogeneous Multi-graph Spatiotemporal Fusion for Long-term Vehicle Trajectory Forecasting

CHEN Zheng, ZHANG Jing, CHEN Bowen, LI Chunyu, GUO Fengxiang, WEI Fuxing*   

  1. Faculty of Transportation Engineering, Kunming University of Science and Technology, Kunming 650500, China
  • Received:2025-05-06 Revised:2025-06-22 Accepted:2025-07-07 Online:2025-08-25 Published:2025-08-25
  • Supported by:
    National Natural Science Foundation of China(52272395)；Yunnan Fundamental Research Projects (202401AS070118)。

PDF

摘要： 车辆轨迹预测的准确率直接影响行车安全性。传统方法仅考虑车辆的运动特征数据，未能充分利用道路环境信息。此外，现有方法在长时域轨迹预测任务中常面临梯度消失等问题，导致预测精度显著下降。为有效解决这些问题，本文提出一种异构多图时空融合的长时域车辆轨迹预测方法。首先，将历史交通信息解耦为道路环境信息与车辆交互信息，并通过图拓扑结构建模方法将上述两类信息分别建模为道路环境图与车辆交互信息图；其次，采用图注意力网络分别对道路环境图与车辆交互信息图进行卷积和池化操作，以获取各自的时空依赖信息；然后，引入门控融合机制，动态调节环境约束与交互行为的贡献权重，得到融合特征序列；最后，通过Mamba网络对融合特征序列解码，输出长时域预测轨迹。仿真结果表明：在5s的预测时域上，本文模型相较于最优基线算法，预测轨迹的平均误差降低22.8%，终点误差降低32.6%，均方根误差降低了18.9%，显著提升了长时域预测精度。

关键词: 智能交通, 轨迹预测, Mamba, 时空轨迹, 门控时空融合

Abstract: The accuracy of vehicle trajectory prediction impacts driving safety significantly. However, conventional methods focus on the features of vehicle kinematic predominantly, while neglecting the comprehensive utilization of road environmental information. Furthermore, the existing approaches often encounter gradient vanishing issues in long-term trajectory prediction tasks, leading to a substantial performance degradation. To address these challenges, this study proposes a novel framework, Heterogeneous Multi-Graph Spatiotemporal Fusion, for Long-Term Vehicle Trajectory Forecasting. Initially, historical traffic data is decoupled into road environmental features and vehicle interaction patterns, which are subsequently modeled as distinct graph topologies-an environmental graph and an interaction graph. Subsequently, graph attention networks are employed to perform convolutional and pooling operations on each graph topology, capturing their spatiotemporal dependencies effectively. Then a gated fusion mechanism is introduced to adjust the contribution weights of environmental constraints and interactive behaviors dynamically by generating optimized fusion features. Finally, the integrated feature sequences are decoded through a Mamba network to produce long-term trajectory predictions. The simulation results demonstrate that, over a 5-second prediction horizon, the proposed model achieves a 22.8% reduction in average error, and a 32.6% decrease in terminal error, and an 18.9% reduction in root mean square error compared to the optimal baseline algorithm, which improve the long-term prediction accuracy significantly.

Key words: intelligent transportation, trajectory prediction, Mamba, spatiotemporal trajectory, gated spatiotemporal fusion

中图分类号: