交通运输系统工程与信息 ›› 2024, Vol. 24 ›› Issue (4): 60-68.DOI: 10.16097/j.cnki.1009-6744.2024.04.007

• 智能交通系统与信息技术 • 上一篇    下一篇

考虑动态交互作用的智能车辆轨迹预测

温惠英,张昕怡,黄俊达,许鹏鹏*   

  1. 华南理工大学,土木与交通学院,广州510641
  • 收稿日期:2024-04-23 修回日期:2024-06-14 接受日期:2024-06-17 出版日期:2024-08-25 发布日期:2024-08-21
  • 作者简介:温惠英(1965- ),女,江西于都人,教授,博士。
  • 基金资助:
    国家自然科学基金 (52372329, 52302433);广东省自然科学基金 (2024A1515011578)。

Intelligent Vehicle Trajectory Prediction Considering Dynamic Interactions

WENHuiying,ZHANG Xinyi,HUANG Junda,XU Pengpeng*   

  1. School of Civil Engineering and Transportation, South China University of Technology, Guangzhou 510641, China
  • Received:2024-04-23 Revised:2024-06-14 Accepted:2024-06-17 Online:2024-08-25 Published:2024-08-21
  • Supported by:
    National Natural Science Foundation of China (52372329, 52302433);Natural Science Foundation of Guangdong Province, China (2024A1515011578)。

摘要: 在多车交互的动态场景中,智能车辆需要具备对周围车辆未来轨迹的预测能力,以实现安全高效行驶。本文提出一种考虑邻车动态交互作用的轨迹预测方法。首先基于目标车辆及周围车辆的历史轨迹信息,构建动态时空关联图,作为交互特征提取模块的输入,再运用图注意力机制获取历史时域上可变的交互特征参数;其次,将目标车辆历史时域信息与可变的交互特征参数融合,嵌入时间注意力机制得到上下文向量,再通过长短时记忆神经网络解码输出目标车辆的未来轨迹;最后,运用CitySim数据集对本文模型进行训练及验证,又采用CQSkyEye数据集对模型进行迁移性实验。结果显示:模型在5s的预测时域上均方根误差为0.82m,与其他预测模型的最优结果(0.96 m)相比,精度提升15%,并且可以提前2s对车辆轨迹进行准确预测;对于迁移性能,本文模型相比其他模型有一定优势,在改变图构建的距离阈值参数后,5s预测时域上的均方根误差为6.43m,对比其他模型最优结果(12.41m),精度提升48%。

关键词: 智能交通, 轨迹预测, 图注意力, 动态交互, 长短时记忆网络

Abstract: For dynamic scenarios involving interaction among multiple vehicles, intelligent vehicles should be able to predict the future trajectories of surrounding vehicles for safe and efficient driving. This paper proposes a trajectory prediction method that considers dynamic interactions among vehicles. First, based on the historical trajectory information of the target and surrounding vehicles, a dynamic spatio-temporal correlation graph is constructed as the input for the interaction feature extraction module. The graph attention mechanism is then used to capture the temporally varying interaction feature parameters. Second, the historical temporal information of the target vehicle is fused with the variable interaction feature parameters. A context vector is obtained by an LSTM encoder embedded with a temporal attention mechanism, followed by using the LSTM decoder to output the future trajectory of the target vehicle. Finally, the proposed model is trained and validated on the CitySim dataset, and transfer experiments are conducted using the CQSkyEye dataset. The results show that the model achieves an RMSE of 0.82 m in a 5 s prediction horizon, demonstrating a 15% improvement in accuracy compared to other popular models. The model also demonstrates the ability to make predictions with less than 2 s lead time. In terms of transferability, the proposed model outperforms others with an RMSE of 6.43 m in the 5 s prediction horizon after adjusting the distance threshold parameter for graph construction, showing an improvement of over 48% in transfer prediction capability.

Key words: intelligent transportation, trajectory prediction, graph attention, dynamic interaction, long short-term memory network

中图分类号: