交通运输系统工程与信息

交通运输系统工程与信息 ›› 2026, Vol. 26 ›› Issue (1): 104-114.DOI: 10.16097/j.cnki.1009-6744.2026.01.010

• 智能交通系统与信息技术 • 上一篇    下一篇

融合时序感知与边界损失的车辆轨迹重构方法

李熙莹1,2，陈泽1,2，李锦3，刘静宇3，盘婳燕3，江倩殷*4   

  1. 1. 中山大学，智能工程学院，广东深圳518107；2.广东省智能交通系统重点实验室，广东深圳518107；3. 广东省公安厅科技信息化总队，广州510050；4.广州航海学院，人工智能学院，广州510725
  • 收稿日期:2025-10-24 修回日期:2025-12-11 接受日期:2025-12-23 出版日期:2026-02-25 发布日期:2026-02-15
  • 作者简介:李熙莹（1972—），女，陕西西安人，教授，博士。
  • 基金资助:
    国家自然科学基金(U21B2090）；由2024YY27项目资助。

Vehicle Trajectory Reconstruction Method Considering Temporal Awareness and Boundary Loss

LI Xiying1,2, CHEN Ze1,2, LI Jin3, LIU Jingyu3, PAN Huayan3, JIANG Qianyin*4   

  1. 1. School of Intelligent Systems Engineering, Shenzhen Campus of Sun Yat-sen University, Shenzhen 518107, Guangdong, China; 2. Guangdong Provincial Key Laboratory of Intelligent Transportation System, Shenzhen 518107, Guangdong, China; 3. Science, Technology and Informatization Corps of Guangdong Provincial Public Security Department, Guangzhou 510050, China; 4. School of Artificial Intelligence, Guangzhou Maritime University, Guangzhou 510725, China
  • Received:2025-10-24 Revised:2025-12-11 Accepted:2025-12-23 Online:2026-02-25 Published:2026-02-15
  • Supported by:
    National Natural Science Foundation of China (U21B2090）；Supported by Project 2024YY27。

PDF

摘要： 自动车辆识别（Automatic Vehicle Identification, AVI）轨迹数据是构建智能交通系统的重要数据基础，但实际应用中常受限于错检和漏检造成的轨迹缺失问题。针对现有轨迹重构方法难以关注全局时间背景与局部边界信息的问题，本文提出一种融合时序感知与边界损失的车辆轨迹重构方法。该方法以Transformer编码器-解码器架构为主体框架，首先，设计时间嵌入模块，融合车辆轨迹序列结构与全局时间背景信息，形成统一向量表示；然后，利用编码器-解码器捕捉轨迹序列的深层依赖关系，并以自回归方式生成重构结果；最后，引入双向边界损失并与标准解码损失进行联合优化，强化模型对缺失轨迹边界强约束信息的关注。在近42万条真实车辆轨迹数据上的实验结果表明，本文方法在缺失率为10%、30%和50%这3种情况下，轨迹重构准确率分别达到95.57%、93.62%和86.36%，各项重构性能指标均优于多种对比基线方法。研究结果表明，全局时间背景感知与局部边界信息约束相结合的策略，能提升深度学习模型在车辆轨迹重构任务中的性能，提高智能交通系统数据完整性。

关键词: 智能交通, 车辆轨迹重构, Transformer模型, 自动车辆识别数据, 时间嵌入

Abstract: Automatic Vehicle Identification (AVI) trajectory data serves as a crucial data foundation for building intelligent transportation systems. However, in practical applications, it is often hindered by the problem of trajectory omission caused by false detections and missed detections. The existing trajectory reconstruction methods are difficult to capture the global time background and local boundary information, this paper proposes a vehicle trajectory reconstruction method combining timing perception and boundary loss. This method utilizes a Transformer encoder-decoder architecture as its core framework. A time embedding module is introduced to fuse the vehicle trajectory sequence structure and global time background information to form a unified vector representation. Then, the encoder-decoder framework is used to describe the deep dependencies of the trajectory sequence, and the reconstruction results are generated in an autoregressive manner. The bidirectional boundary loss was then jointly optimized with the standard decoding loss to strengthen the model's attention to the strong constraint information of the missing trajectory boundary. Experimental results based on nearly 420 000 real vehicle trajectory data show that the method achieves trajectory reconstruction accuracies of 95.57% , 93.62%, and 86.36% under missing rates of 10%, 30%, and 50%, respectively, with all reconstruction performance indicators outperforming various baseline methods. The results shows that the proposed methods can improve the performance of deep learning models in vehicle trajectory reconstruction tasks and improve the data integrity of intelligent transportation systems.

Key words: intelligent transportation, vehicle trajectory reconstruction, Transformer model, automatic vehicle identification data, time embedding

中图分类号: