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 (3): 107-116.DOI: 10.16097/j.cnki.1009-6744.2025.03.010

• 智能交通系统与信息技术 • 上一篇    下一篇

考虑时空特征动态耦合的车辆轨迹预测方法

高远，付金龙，冯文文*   

  1. 东北林业大学，土木与交通学院，哈尔滨150040
  • 收稿日期:2025-01-26 修回日期:2025-03-06 接受日期:2025-03-13 出版日期:2025-06-25 发布日期:2025-06-20
  • 作者简介:高远(1993—)，男，山东菏泽人，副教授，博士。
  • 基金资助:
    中央高校基本科研业务费专项资金(2572022BG01)；黑龙江省哲学社会科学研究规划项目(24GLC014)。

Vehicle Trajectory Prediction Method Considering Dynamic Coupling of Spatial-temporal Features

GAO Yuan, FU Jinlong, FENG Wenwen*   

  1. School of Civil and Transportation Engineering, Northeast Forestry University, Harbin 150040, China
  • Received:2025-01-26 Revised:2025-03-06 Accepted:2025-03-13 Online:2025-06-25 Published:2025-06-20
  • Supported by:
    Fundamental Research Funds for the Central Universities of Ministry of Education of China (2572022BG01)；Philosophy and Social Science Foundation of Heilongjiang Province (24GLC014)。

PDF

摘要： 在复杂的多车动态交互场景中，智能车辆需要精准感知并预测周围车辆的行驶轨迹，以确保安全高效的行驶。针对现有模型未能充分考虑多维特征之间动态耦合关系的问题，本文提出一种基于时空交叉注意力机制的车辆轨迹预测方法。首先，采用空间注意力模块，从目标车辆及周围车辆的历史轨迹数据中提取车辆间的动态交互特征；其次，将动态交互特征参数输入至长短时记忆神经网络编码器，从时域角度捕捉跨时间依赖关系；然后，将编码器的隐藏状态输入至结合了傅里叶变换和学习性路由器的交叉注意力模块，在频域上捕捉跨时间依赖关系，并进一步提取多维特征间的耦合特性；之后，通过长短时记忆神经网络解码器生成目标车辆的未来轨迹，采用NGSIM数据集对模型进行训练、验证和测试。结果显示，模型在5s的预测时域上均方根误差为0.74m，与其他预测模型的最优结果(0.82m)相比，精度提升了10%。

关键词: 智能交通, 轨迹预测, 交叉注意力机制, 特征耦合, 长短时记忆网络

Abstract: In complex multi-vehicle dynamic interaction scenarios, intelligent vehicles need to accurately perceive and predict the driving trajectories of surrounding vehicles to ensure safe and efficient driving. To address the issue that existing models fail to fully consider the dynamic coupling relationships among multi-dimensional features, this paper proposes a vehicle trajectory prediction method based on a spatio-temporal cross-attention mechanism. First, a spatial attention module is adopted to extract the dynamic interaction features between the target vehicle and surrounding vehicles from their historical trajectory data. Then, the obtained dynamic interaction feature parameters are input into a long short-term memory (LSTM) neural network encoder to capture cross-time dependencies from the time domain perspective. Subsequently, the hidden state of the encoder is input into a cross-attention module that combines Fourier transform and a learnable router to capture cross-time dependencies in the frequency domain and further extract the coupling features among multi-dimensional features. At last, the future trajectory of the target vehicle is generated through an LSTM neural network decoder. The model is trained, validated, and tested using the next generation simulation (NGSIM) dataset. The results show that the model has a root mean square error of 0.74 meters in the 5-second prediction time domain, which is a 10% improvement in accuracy compared to the best results of other prediction models (0.82 meters).

Key words: intelligent transportation, trajectory prediction, cross-attention mechanism, feature coupling, long short-term memory network

中图分类号: