欢迎访问《汽车安全与节能学报》，
汽车安全与节能学报

汽车安全与节能学报 ›› 2025, Vol. 16 ›› Issue (5): 688-697.DOI: 10.3969/j.issn.1674-8484.2025.05.003

• 汽车安全 • 上一篇    下一篇

基于混合神经网络的交织区危险驾驶与风格的识别

程泽阳1(), 段奕阳1, 杨蒙蒙2,*(), 冯忠祥1, 王鹤3, 朱晓俊3, 保丽霞4   

  1. 1.合肥工业大学 汽车与交通工程学院，安徽，合肥，230009，中国
    2.清华大学 车辆与运载学院，北京，100084，中国
    3.江淮前沿技术协同创新中心，合肥，230088，中国
    4.上海市城市建设设计研究总院（集团）有限公司，上海，200082，中国
  • 收稿日期:2025-01-28 修回日期:2025-06-04 出版日期:2025-10-31 发布日期:2025-11-10
  • 通讯作者: *杨蒙蒙，助理研究员。E-mail：yangmm_qh@tsinghua.edu.cn。
  • 作者简介:程泽阳（1990—），男（汉），副研究员。E-mail：chengzeyang@hfut.edu.cn。
  • 基金资助:
    国家自然科学基金青年项目(52202411);智能绿色车辆与交通全国重点实验室开放基金课题(KFY2418);江淮前沿技术协同创新中心追梦基金(2023-ZM01G006)

Recognition of the dangerous driving behaviors and the driving styles in weaving areas based on a hybrid neural network

CHENG Zeyang1(), DUAN Yiyang1, YANG Mengmeng2,*(), FENG Zhongxiang1, WANG He3, ZHU Xiaojun3, BAO Lixia4   

  1. 1. School of Automotive and Transportation Engineering, Hefei University of Technology, Hefei 230009, China
    2. School of Vehicle and Mobility, Tsinghua University, Beijing 100084, China
    3. Jianghuai Advance Technology Center, Hefei 230088, China
    4. Shanghai Urban Construction Design and Research Institute, Shanghai 200082, China
  • Received:2025-01-28 Revised:2025-06-04 Online:2025-10-31 Published:2025-11-10

RichHTML

94

PDF

56

摘要： 为辨识和预测交织区危险驾驶行为及驾驶风格，提出了一种基于历史轨迹数据的混合神经网络分析方法。利用k-means++算法，对不同危险驾驶行为进行聚类分析，提取纵向速度、横向与纵向加速度的关键特征，以表征不同类型的驾驶风格；基于长短期记忆网络（LSTM） -卷积神经网络（CNN） -知识注意机制（KAN），构建驾驶风格预测模型；进行了仿真与消融对比实验。结果显示：该模型的受试者操作特征（ROC）曲线下面积（AUC）为0.846；相较于使用LSTM、CNN-LSTM、LSTM-KAN的模型，该模型危险驾驶行为及驾驶风格分类预测的精度提升了6.73%、3.12%、4.72%; 泛化性验证精度分别提升了6.3%、2.5%、3.9%。

关键词: 智能交通, 城市快速路交织区, 危险驾驶行为, 驾驶风格, k-means++算法, 聚类分析, 深度学习, 混合神经网络

Abstract:

A hybrid neural network analysis method was proposed based on historical trajectory data to identify and predict dangerous driving behaviors and driving styles in weaving areas. The different dangerous driving behaviors were clustered and analyzed by using the K-means++ algorithm with the key features of the longitudinal velocity, the lateral acceleration, and the longitudinal acceleration being extracted to characterize different driving styles. A driving style prediction model was constructed based on a Long Short-Term Memory (LSTM) network, a Convolutional Neural Network (CNN), and a Knowledge-Attention Network (KAN), with conducting digital simulations and ablation comparison experiments. The results show that the model has the Area Under Curve (AUC) of the Receiver Operating Characteristic (ROC), a dimension-one quantity, of 0.846. the model's classification and prediction accuracy of dangerous driving behaviors and driving styles increased by 6.73%, 3.12%, and 4.72%, while increasing the generalization verification accuracy by 6.3%, 2.5%, and 3.9%, compared with models using LSTM, CNN-LSTM, and LSTM-KAN.

Key words: Intelligent transport, urban expressway weaving area, dangerous driving behaviors, driving styles, k-means++ algorithm, cluster analysis, deep learning, hybrid neural

中图分类号: