多车道高速合流区车辆换道识别及轨迹分析  

Lane Changing Identification and Trajectory Analysis in Confluence Area of Multi-lane Expressway

作　　者：齐龙 郝艳军 徐婷[3] 陈亦新 王孝冬 QI Long;HAO Yan-jun;XU Ting;CHEN Yi-xin;WANG Xiao-dong(School of Automotive Engineering,Shandong Jiaotong University,Jinan,Shandong 250357,China;Shanxi Intelligent Transportation Institute Co.,Ltd.,Taiyuan,Shanxi 030006,China;School of Transportation Engineering,Chang’an University,Xi’an,Shaanxi 710064,China;Guangzhou Yuelu Traffic Technology Co.,Ltd.,Guangzhou,Guangdong 510630,China)

机构地区：[1]山东交通学院汽车工程学院,山东济南250357 [2]山西省智慧交通研究院有限公司,山西太原030006 [3]长安大学运输工程学院,陕西西安710064 [4]广州越路交通科技有限公司,广东广州510630

出　　处：《公路交通科技》2024年第2期173-181,共9页Journal of Highway and Transportation Research and Development

基　　金：国家自然科学基金项目(51878066);陕西省自然科学基础研究计划项目(2021JQ-276);山东交通学院科研基金项目(Z201901)。

摘　　要：为了提高多车道高速公路合流交织区交通安全,研究车辆换道轨迹特征并识别换道意图。首先,利用无人机采集目标区域为单向两条车道以上的高速公路合流区域车辆自然状态下运行轨迹视频,经Tracker运动学轨迹软件提取每条轨迹的时间和位置等信息,获取车辆行驶轨迹,累计获得自然驾驶轨迹数,并且根据车头时距、车速、加速度数据得到跟驰和换道片段,并对两类驾驶行为的时长和距离分布进行了分析,得到车辆换道轨迹片段数据。其次,结合车辆速度、加速度和变加速度,通过K-means++聚类将驾驶风格分为“常规型”、“激进型”和“保守型”3种,确定了用随机森林模型对不同风格的驾驶人轨迹进行了识别,进而选取基于XGBoost,LightGBM两者的Stacking融合模型对车辆换道意图进行了识别。最后,构建机器学习CNN-LSTM模型进行轨迹的预测模型。结果表明:采用K-means++将驾驶风格聚类为3类模型的综合效果最优,选定其聚类结果为轨迹片段驾驶风格标签值,随机森林的准确率较好,选用Stacking融合模型的准确率适用于驾驶轨迹识别,从换道轨迹的预测准确角度判定R2处于0.62水平,并且当时间窗为2 s时,模型对预测换道轨迹能做出较为准确的预测;研究实现了通过驾驶行为识别预测轨迹,为实现车辆的实时碰撞风险识别应用提供理论基础,同时可以优化自动驾驶车辆的行驶轨迹,提高高速公路入口复杂交通流状态下的安全性。In order to improve the traffic safety in the expressway confluence area,the characteristics of vehicle lane changing trajectory are studied and the intention of lane changing is identified.Firstly,the UAV is used to capture the running track video of vehicles in the expressway confluence area under natural state.The time and position information of each track are extracted with Tracker kinematics track software,the vehicle running track is obtained,and the number of natural driving trajectories is accumulated.The segments of car-following and lane-changing are obtained according to the data of time distance,speed and acceleration,and the time and distance distributions of 2 types of driving behaviors are analyzed.Secondly,combining with vehicle speed,acceleration and variable acceleration,K-means++clustering is used to classify the driving styles into‘conventional’,‘radical’and‘conservative’.Stochastic forest model is determined to identify the trajectories of drivers with different styles,and then Stacking fusion model based on XGBoost and LightGBM is selected to identify the intention of vehicle lane changing.Finally,a machine learning CNN-LSTM model is constructed to predict the trajectories.The result shows that(1)the comprehensive effect of clustering driving styles into 3 categories by using K-means++is the best,the clustering result is selected as the label value of driving style in trajectory segments,and the accuracy of random forest is good;(2)the accuracy of using the Stacking fusion model is suitable for driving trajectory recognition,R2 is determined to be at the level of 0.62 from the perspective of lane changing trajectory prediction accuracy,when the time window is 2 s,the model can make a more accurate prediction on lane changing trajectory prediction;(3)the study has realized the purpose of predicting the trajectory through driving behavior recognition,providing an application basis for realizing real-time collision risk identification of vehicles.At the same time,it can optimize th

关 键 词：交通工程 交通系统安全 聚类分析 换道行为 合流区 

分 类 号：U491.234[交通运输工程—交通运输规划与管理]