汽车工程 ›› 2024, Vol. 46 ›› Issue (5): 766-775.doi: 10.19562/j.chinasae.qcgc.2024.05.003

• • 上一篇    

考虑风险时空分布特征的跟驰工况行车风险预测

宋东鉴1,赵健1,朱冰1,佟静2(),韩嘉懿1,刘斌3   

  1. 1.吉林大学,汽车仿真与控制国家重点实验室,长春 130022
    2.吉林大学汽车工程学院,长春 130022
    3.中国第一汽车集团有限公司,长春 130011
  • 收稿日期:2023-05-25 修回日期:2023-07-12 出版日期:2024-05-25 发布日期:2024-05-17
  • 通讯作者: 佟静 E-mail:tongjing@jlu.edu.cn
  • 基金资助:
    国家自然科学基金(52172386);吉林省自然科学基金(20210101057JC)

Driving Risk Prediction Under Car-Following Conditions Considering Risk Spatiotemporal Distribution Characteristics

Dongjian Song1,Jian Zhao1,Bing Zhu1,Jing Tong2(),Jiayi Han1,Bin Liu3   

  1. 1.Jilin University,State Key Laboratory of Automotive Simulation and Control,Changchun 130022
    2.College Automotive Engineering,Jilin University,Changchun 130022
    3.China FAW Group Co. ,Ltd. ,Changchun 130011
  • Received:2023-05-25 Revised:2023-07-12 Online:2024-05-25 Published:2024-05-17
  • Contact: Jing Tong E-mail:tongjing@jlu.edu.cn

摘要:

行车风险预测对提升智能汽车行车安全性至关重要,为此本文提出了一种跟驰工况下的行车风险预测模型(car-following risk prediction model, CRPM)。跟驰中车辆的减速能够反映驾驶人的认知风险,故根据车辆纵向加速度标注跟驰风险等级,并构建基于各向异性行车风险场的风险时空分布特征以用作CRPM的输入。CRPM通过卷积神经网络提取风险的空间分布特性,利用双向长短期记忆网络和注意力机制处理风险的时序依赖关系,最终输出跟驰风险等级。CRPM在航拍数据集AD4CHE上进行训练和测试。结果表明,CRPM具有良好的预测精度和提前预测时间,预测准确率达99.67%,在风险发生前2 s预测准确率为96.73%。

关键词: 行车风险预测, 风险时空分布, 跟驰, 航拍数据

Abstract:

Risk prediction of driving is very important for improving the driving safety of intelligent vehicles. In this paper, a car-following risk prediction model (CRPM) is proposed. The deceleration of the vehicle in a car-following process can reflect the driver's cognitive risk, so longitudinal accelerations are taken as the basis for the car-following risk level labeling, and the risk spatiotemporal distribution characteristics based on the anisotropic driving risk field are constructed and used as input of the CRPM. The CRPM extracts the spatial distribution characteristics through Convolutional Neural Network, and processes the temporal dependence relationship by bidirectional Long Short-Term Memory network and attention mechanism, and finally outputs the car-following risk level. The CRPM is trained and tested on the aerial driving dataset AD4CHE. The results show that the CRPM has good prediction accuracy and long advanced prediction time, with a prediction accuracy of 99.67%, and a prediction accuracy of 96.73% within 2 s before the occurrence of risk.

Key words: driving risk prediction, risk spatiotemporal distribution, car-following, aerial data