交通运输系统工程与信息 ›› 2025, Vol. 25 ›› Issue (4): 116-125.DOI: 10.16097/j.cnki.1009-6744.2025.04.012

• 智能交通系统与信息技术 • 上一篇    下一篇

基于物理信息深度学习的交叉口车辆轨迹补全方法

郑立勇1,2 ,孙剑1 ,饶红玉*2 ,邵健轩2 ,赵威2 ,郝勇刚2   

  1. 1. 同济大学,道路与交通工程教育部重点实验室,上海201804;2.杭州海康威视数字技术股份有限公司,上海201203
  • 收稿日期:2025-03-31 修回日期:2025-05-20 接受日期:2025-05-21 出版日期:2025-08-25 发布日期:2025-08-25
  • 作者简介:郑立勇(1992—),男,江西上饶人,博士生。
  • 基金资助:
    上海市科技创新行动计划(23DZ1203400);上海市科委项目 (22dz1203200)。

Vehicle Trajectory Imputation at Intersection Based on Physics-informed Deep Learning

ZHENG Liyong1,2, SUN Jian1, RAO Hongyu*2, SHAO Jianxuan2, ZHAO Wei2, HAO Yonggang2   

  1. 1. The Key Laboratory of Road and Traffic Engineering, Ministry of Education, Tongji University, Shanghai 201804, China; 2. Hangzhou Hikvision Digital Technology Co Ltd, Shanghai 201203, China
  • Received:2025-03-31 Revised:2025-05-20 Accepted:2025-05-21 Online:2025-08-25 Published:2025-08-25
  • Supported by:
    Shanghai Action Plan for Science, Technology and Innovation (23DZ1203400);Science and Technology Commission of Shanghai Municipality (22dz1203200)。

摘要: 车辆轨迹数据在智能交通系统中有着诸多应用,但其实际应用效果常常受数据缺失问题影响。雷达和视频融合感知技术的迅速发展虽然实现了车辆轨迹数据的全天候采集,但在交叉口场景中仍然面临雷达对排队静止目标不敏感,大型车辆遮挡等原因导致数据缺失问题。针对交叉口车辆轨迹数据缺失,本文提出一种基于物理信息深度学习的补全算法(Transformer-Full-Velocity-Difference, TF-FVD),将FVD跟驰模型的监督信号引入到Transformer模型的训练过程中,并增加信号灯状态编码模块以考虑交通信号约束。基于雷视轨迹数据集的实验结果表明:FVD模型监督信号和信号灯状态编码模块的引入分别带来了11.6%和15.6%的精度提升;在SinD(Signalized INtersection Dataset)公开数据集中,本文提出的TF-FVD模型较纯数据驱动SOTA(StateoftheArt)算法精度提升了25.3%;基于补全轨迹计算的车辆延误时间分布误差降低了9.14%,体现了其在实际应用中的价值。

关键词: 交通工程, 车辆轨迹补全, 物理信息深度学习, 雷视融合感知

Abstract: The trajectory data of vehicles has numerous applications in intelligent transportation systems (ITS). However, their practical effectiveness is often hampered by data missing. Although the rapid development of radar-and-video-fused perception technology has enabled all-day collection of vehicle trajectory data, some challenges still persist in intersection scenarios, such as the insensitivity of radar to stationary targets and the occlusion by large vehicles. To address the missing data of vehicle trajectory in intersections, this paper proposed a novel completion algorithm (Transformer-Full-Velocity-Difference, TF-FVD) based on physics-informed deep learning, which incorporates the supervision signal of the FVD car-following model into the training process of the Transformer deep learning model, and adds a traffic light state encoding module to account for the traffic rule constraints on vehicle movement. The experimental results based on the radar-video-fused trajectory dataset show that the introduction of the FVD model and the traffic light state encoding module led to the improvements of accuracy by 11.6% and 15.6% respectively. In public SinD (Signalized INtersection Dataset) dataset, the proposed TF-FVD model achieved a 25.3% accuracy improvement compared to the SOTA (State of the Art) data-driven algorithm. The distribution error of travel delay calculated from the imputed trajectories decreased by 9.14%, which implies its value in following applications.

Key words: traffic engineering, vehicle trajectory imputation, physics-informed deep learning, radar-and-video-fused perception

中图分类号: