交通运输系统工程与信息

交通运输系统工程与信息 ›› 2024, Vol. 24 ›› Issue (2): 63-75.DOI: 10.16097/j.cnki.1009-6744.2024.02.007

• 智能交通系统与信息技术 • 上一篇    下一篇

考虑多模态数据的重载货车危险驾驶行为识别方法

吴建清1，2 ，张子毅1，2 ，王钰博3 ，张昱4 ，田源*1，2   

  1. 1. 山东大学，齐鲁交通学院，济南250061；2.山东大学，苏州研究院，江苏苏州215123； 3. 济南城建集团有限公司，济南250031；4.山东省交通科学研究院，济南250012
  • 收稿日期:2024-01-18 修回日期:2024-02-09 接受日期:2024-02-26 出版日期:2024-04-25 发布日期:2024-04-25
  • 作者简介:吴建清(1988- )，男，山东烟台人，教授，博士。
  • 基金资助:
    国家重点研发计划(2022YFB2602102)；国家自然科学基金(52002224)。

Method for Identifying Dangerous Driving Behaviors in Heavy-duty Trucks Based on Multi-modal Data

WUJianqing1,2,ZHANG Ziyi1,2,WANG Yubo3,ZHANGYu4,TIANYuan*1,2   

  1. 1. School of Qilu Transportation, Shandong University, Jinan 250061, China; 2. Suzhou Research Institute, Shandong University, Suzhou 215123, Jiangsu, China; 3. Jinan Urban Construction Group Co Ltd, Jinan 250031, China; 4. Shandong Transportation Institute, Jinan 250012, China
  • Received:2024-01-18 Revised:2024-02-09 Accepted:2024-02-26 Online:2024-04-25 Published:2024-04-25
  • Supported by:
    NationalKeyResearchandDevelopmentProgram of China (2022YFB2602102)；National Natural Science Foundation of China (52002224)。

PDF

摘要： 综合考虑货车操纵数据、驾驶员眼动数据和心电数据，本文提出一种多模态的重载货车危险驾驶指标构建和行为识别方法。首先，设计自然驾驶实验，利用车辆惯导、眼动仪及心理数据记录仪这3种设备采集车辆运行、眼动及心电等多模态驾驶数据，通过多设备时间同步及数据清洗，构建多模态驾驶数据集。其次，将重载货车危险驾驶行为分为危险驾驶操纵行为和疲劳驾驶行为两类，通过提取数据特征，构建9种危险驾驶行为指标，表征超速、速度不稳、急变速、急换道及疲劳驾驶这5种危险驾驶行为。针对危险驾驶操纵行为，组合文献调研、指标计算及四分位差法确定危险行为特征参数阈值；针对疲劳驾驶行为，通过因子分析和K均值聚类法划分疲劳驾驶类型。最后，构建重载货车危险驾驶行为数据集，采用随机森林分类模型识别危险驾驶行为，并与BP神经网络、K近邻及支持向量机等分类模型对比。结果表明，随机森林模型对于5种危险驾驶行为的分类准确率均在90%以上，整体优于其他分类算法，能够较高精度地识别驾驶中出现的危险驾驶行为。本文的多模态重载货车危险驾驶指标构建和分类方法能够用于危险驾驶行为识别，为驾驶员多模态危险驾驶行为预警提供思路和理论依据。

关键词: 交通工程, 驾驶行为识别, 阈值量化, 重载货车, 随机森林

Abstract: This paper proposes a multi-modal method to identify dangerous driving behaviors of heavy-duty trucks, which integrates driving operation data, eye-tracking data and electrocardiogram (ECG) data in the analysis. A naturalistic driving experiment is designed to collect driving data using three types of devices: vehicle inertial navigation systems, eye-tracking decoders, and physiological data recorders. A multi-modal driving dataset is established through data synchronizing and data cleaning processes. The dangerous driving behaviors are divided into two categories: dangerous manipulation behaviors and fatigue driving behaviors. By extracting data features, nine dangerous driving behavior indicators are defined to represent five types of dangerous driving behaviors, including speeding, unstable speed, rapid speed changing, rapid lane changing, and fatigue driving. For dangerous manipulation behaviors, characteristic thresholds are determined through literature review, indicator calculation, and interquartile range method. For fatigue driving behaviors, fatigue driving levels are identified through factor analysis and K-means clustering methods. A random forest (RF) classification model is then developed to identify dangerous driving behavior. When compared to traditional methods, including back propagation neural network (BPNN), K-nearest neighbors (KNN), support vector machine (SVM), the proposed model surpassed others in terms of accuracy and fitting performance. The model achieved a classification accuracy of over 90% for all types of dangerous driving behaviors. The results prove that the proposed methods are effective in identifying dangerous driving behaviors and it provides a theoretical basis for multimodal warning systems of dangerous driving behaviors.

Key words: traffic engineering, driving behavior recognition, threshold quantization, heavy-duty trucks, random forest

中图分类号: