中国安全科学学报 ›› 2024, Vol. 34 ›› Issue (6): 164-172.doi: 10.16265/j.cnki.issn1003-3033.2024.06.0720

• 安全工程技术 • 上一篇    下一篇

基于EKF-GRU的车辆轨迹预测

张传莹1(), 徐国艳1, 陈志发1, 周彬1, 陈立伟2, 洪玮2   

  1. 1 北京航空航天大学 交通科学与工程学院,北京 100191
    2 贵州开源爆破工程有限公司,贵州 贵阳 551400
  • 收稿日期:2023-12-14 修回日期:2024-03-19 出版日期:2024-06-28
  • 作者简介:

    张传莹 (1997—),男,山东潍坊人,硕士研究生,主要研究方向为轨迹预测和碰撞风险评估。E-mail:

    徐国艳 副教授

  • 基金资助:
    国家重点研发计划课题项目(2022YFB4703702)

Vehicle trajectory prediction based on EKF-GRU

ZHANG Chuanying1(), XU Guoyan1, CHEN Zhifa1, ZHOU Bin1, CHEN Liwei2, HONG Wei2   

  1. 1 School of Transportation Science and Engineering, Beihang University, Beijing 100191, China
    2 Guizhou Kaiyuan Explosive Engineering Co., Ltd., Guiyang Guizhou 551400, China
  • Received:2023-12-14 Revised:2024-03-19 Published:2024-06-28

摘要:

为提升行车安全,实现自动驾驶车辆正确的决策规划,提出基于扩展卡尔曼滤波(EKF)-门控循环单元(GRU)的车辆轨迹预测方法,结合学习方法与物理模型,在提升预测精度的同时,提高轨迹预测的合理性。首先,基于GRU构建预测网络,通过提取车辆的历史轨迹特征预测车辆的纵向加速度及横摆角速度;其次,基于车辆非线性运动学构建EKF状态估计器,结合观测值生成车辆未来有限时域的行驶轨迹;最后,在高速公路多车轨迹数据集NGSIM I-80和US-101上进行轨迹预测方法验证。结果表明:采用传统的物理模型生成预测轨迹,其最终距离误差(FDE)、均方根误差(RMSE)、平均距离误差(ADE)值分别为6.48、7.69和3.03 m。相比之下,利用EKF-GRU生成的预测轨迹表现出更高的准确性,对应的数值分别为5.45、6.67和2.56 m,分别提升15.90%、13.26%和15.51%。

关键词: 扩展卡尔曼滤波(EKF), 门控循环单元(GRU), 车辆轨迹, 轨迹预测, NGSIM数据集, 神经网络

Abstract:

To enhance the driving safety and achieve correct decision planning for autonomous vehicles, a safe driving trajectory prediction method based on EKF-GRU was proposed. By combining learning-based methods with physics-based approaches, the prediction accuracy was improved and the rationality of the predicted trajectories was enhanced. In the first step of this method, a prediction network was constructed based on GRU to predict the longitudinal acceleration and yaw angular velocity of vehicles by extracting historical trajectory features. In the second step, an EKF state estimator was built based on the nonlinear vehicle kinematics to generate the vehicle's future limited-time trajectory, incorporating the observations obtained previously. The trajectory prediction method was validated on the NGSIM I-80 and US-101 multi-vehicle trajectory datasets. Experimental results demonstrate that the final distance errors (FDE), root mean square errors (RMSE), and average distance errors (ADE) of the predicted trajectories generated by traditional physics-based methods are 6.48, 7.69 and 3.03 meters, respectively. In contrast, trajectories predicted using EKF-GRU exhibit higher accuracy, and the corresponding values are 5.45, 6.67 and 2.56 meters, respectively. This represents improvements of 15.90%, 13.26% and 15.51%.

Key words: extended Kalman filtering (EKF), gate recurrent unit (GRU), vehicle trajectory, trajectory prediction, next generation simulation (NGSIM) dataset, neural network

中图分类号: