基于DE-EL的城市快速路合流区危险驾驶行为识别方法  

A Recognition Method for Risky Driving Behaviors of Urban Expressway Merging Area Based on DE-EL Model

作　　者：谢厅 刘星良 刘唐志[1] 徐进[1] XIE Ting;LIU Xingliang;LIU Tangzhi;XU Jin(School of Traffic&Transportation,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学交通运输学院,重庆400074

出　　处：《交通信息与安全》2024年第6期23-30,共8页Journal of Transport Information and Safety

基　　金：国家自然科学基金项目(52172341);重庆市自然科学基金面上项目(CSTB2022NSCQ-MSX0519、CSTB2022NSCQMSX1516)资助。

摘　　要：为提高快速路合流区行车安全水平,实现合流区危险驾驶行为准确识别与交通事故预防,基于车辆轨迹数据提出了1种合流区驾驶人危险驾驶操作行为辨识方法。依托合流区交通航拍视频轨迹数据,运用风险度量法与四分位差法确定4类合流区驾驶人危险驾驶操作行为特征指标阈值。通过前期建立的合流区危险驾驶行为谱,计算驾驶人危险操作得分G,标记危险驾驶人,实现驾驶人分类。选用ROS、SMOTE、ADASYN数据均衡算法(data equalization,DE)对不平衡数据集中的危险驾驶人样本进行扩充,降低轨迹数据集的不平衡度。联合XGBoost、LGBM、AdaBoost集成学习分类算法(ensemble learning,EL)建立DE-EL模型,以车速、变速、横向操作、位置特征以及时间占比5类特征参数变量作为输入,对合流区驾驶人危险驾驶操作行为进行识别。通过Spearman相关性分析对DE-EL识别模型输入特征参数进行优化,提升合流区危险驾驶操作行为识别模型的性能,最终从模型的精确率、召回率、F1值和AUC值确定最优合流区危险驾驶行为识别模型。研究表明:合流区驾驶人行车风险水平与横向操作关联度最高,与车辆速度关联度较低;不平衡的轨迹数据集通过单一的EL算法难以有效识别危险驾驶操作行为,DE算法可显著提升分类算法的性能;特征优化工程后,DE-EL识别模型的性能得到了提升,结果表明SMOTE-LGBM模型对合流区危险驾驶行为的识别效果最好,精确率为93.4%,召回率为92.1%,F1值为0.927,AUC值为0.933,模型可用于合流区危险驾驶行为识别、预警以及干预。A method for recognizing risky driving behaviors using vehicle trajectory data is established to improve safety and prevent traffic accident in urban expressway merging areas.The characteristic thresholds of four types of risky driving behaviors are firstly determined using a risk assessment approach and the interquartile range method.Subsequently,drivers’risk scores(G)are calculated using the established spectrum of risky driving behaviors,en-abling the classification of drivers as safe or risky.To balance the datasets,the driving risk samples are augmented by data equalization(DE)algorithms(ROS,ADASYN,and SMOTE).Combining ensemble learning(EL)algo-rithms(XGBoost,LGBM and AdaBoost)to build various DE-EL models for risky driving behaviors recognition.The Spearman correlation coefficient is used to optimize the input feature parameters,which include five categories:vehicle speed,acceleration and deceleration,lateral operation,position characteristics and time occupation ratio.The optimal recognition model is is determined based on precision rate,recall rate,F1-score and AUC value.The results show that the level of driver risk is most strongly correlated with driver lateral operation and less so with ve-hicle speed in merging areas.The unbalanced trajectory dataset makes it difficult to effectively identify risky driving behaviors by the EL algorithm,while the DE algorithm can improve the properties of the classification algorithm.After optimizing the input feature parameters,the performance of the DE-EL recognition model improves,and the SMOTE-LGBM model is the best one with precision rate of 93.4%,recall rate of 92.1%,F1-score of 0.927,and AUC value of 0.933.This model is applicable for recognizing,warning,intervening in risky driving behaviors in merging areas.

关 键 词：交通安全 合流区驾驶行为 危险驾驶行为谱 集成学习 数据均衡算法 

分 类 号：U491.25[交通运输工程—交通运输规划与管理]