交通运输系统工程与信息

交通运输系统工程与信息 ›› 2025, Vol. 25 ›› Issue (3): 299-307.DOI: 10.16097/j.cnki.1009-6744.2025.03.027

• 系统工程理论与方法 • 上一篇    下一篇

交通事故智能定责深度学习集成方法研究

黄钢*，高岩，赵冬，寿任祯   

  1. 公安部交通管理科学研究所，江苏无锡214151
  • 收稿日期:2025-02-28 修回日期:2025-03-25 接受日期:2025-04-09 出版日期:2025-06-25 发布日期:2025-06-22
  • 作者简介:黄钢(1991—)，男，湖北荆州人，助理研究员。
  • 基金资助:
    国家重点研发计划(2023YFC3009700)。

Deep Learning Ensemble Method for Intelligent Liability Determination in Traffic Accidents

HUANG Gang*, GAO Yan, ZHAO Dong, SHOU Renzhen   

  1. Traffic Management Research Institute of the Ministry of Public Security, Wuxi 214151, Jiangsu, China
  • Received:2025-02-28 Revised:2025-03-25 Accepted:2025-04-09 Online:2025-06-25 Published:2025-06-22
  • Supported by:
    National Key Research and Development Program of China (2023YFC3009700)。

PDF

摘要： 为研究道路交通事故当事人责任智能认定方法，将双方当事人责任结果划分为全责-无责、主责-次责、同等-同等、次责-主责、无责-全责。在基于特征分析和特征工程筛选训练样本特征的基础上，深度离散并独热编码事故特征，结合线性模型和深度神经网络，训练交通事故智能定责深度学习集成模型；对模型的网络结构和超参数进行优化，建立复杂网络结构下的交通事故责任智能认定模型；基于历史交通事故交叉验证模型的可行性。结果表明：深度学习集成方法对交通事故责任认定的准确率有明显提升，对于全责无责和无责-全责类事故，模型准确率达到0.91，其他认定结果的准确率达到0.78；深度学习集成方法对比传统机器学习准确率提高30%以上，对比单一全连接神经网络准确率提高8%以上。深度学习集成方法对交通事故智能定责的边界确定具有明显的改善作用。

关键词: 智能交通, 智能定责, 深度学习, 交通事故, 集成算法, 动态混合模型

Abstract: To investigate intelligent methods for determining the liability of both parties involved in road traffic accidents, the liability outcomes of the involved parties were categorized into five types, full liability-no liability, primary liability-secondary liability, equal liability-equal liability, secondary liability-primary liability, and no liability-full liability. Based on feature analysis and feature engineering to screen training sample characteristics, accident features were thoroughly discretized and one-hot encoded. A deep learning ensemble model for intelligent liability determination in traffic accident was trained by integrating linear models and deep neural networks. The intelligent liability determination model under complex network architectures was established through network structure and hyperparameters. The feasibility of the proposed model was cross-validated using historical traffic accident data. The results show that the deep learning ensemble method significantly elevates the accuracy of traffic accident liability determination. For the full liability-no liability and no liability-full liability accident types, the accuracy of results from the proposed model reaches 0.91, while the accuracies of liability determination results on other accident types are 0.78. The ensemble approach elevates accuracy by over 30% and more than 8% respectively compared to traditional machine learning methods and a single fully connected neural network. The deep learning ensemble method demonstrates notable enhancements in defining liability boundaries for intelligent traffic accident adjudication.

Key words: intelligent transportation, intelligent liability determination, deep learning, traffic accidents, ensemble algorithm; dynamic hybrid ensemble

中图分类号: