基于GS-AdaBoost优化模型的隧道入口车辆换道行为预测  

Prediction of Lane-Changing Behavior of Vehicles at Tunnel Entrances Based on GS-AdaBoost Optimized Model

作　　者：赵荣达 张妍 梁洪健 陈亮[2] 罗方 田蔚楠 秦雅琴[2] ZHAO Rongda;ZHANG Yan;LIANG Hongjian;CHEN Liang;LUO Fang;TIAN Weinan;QIN Yaqin(Yunnan Communications Investment&Construction Group Co.,Ltd.,Kunming 650103,China;Faculty of Transportation Engineering,Kunming University of Science and Technology,Kunming 650500,China)

机构地区：[1]云南省交通投资建设集团有限公司,云南昆明650103 [2]昆明理工大学交通工程学院,云南昆明650500

出　　处：《昆明理工大学学报（自然科学版）》2024年第1期147-155,共9页Journal of Kunming University of Science and Technology(Natural Science)

基　　金：云南交投集团科技研发项目(YCIC-YF-2021-05);云南省大学生创新创业训练计划项目(S202210674132);云南省交通运输厅公路安全与减灾防灾关键技术研究项目(SZKM202231021);云南省交通运输厅科技创新及示范项目(2022-23-3)。

摘　　要：准确估计隧道入口车辆行为对于提高隧道行车安全与效率具有重要意义.现有车辆行为预测较少考虑车辆之间的交互信息,且车辆运动数据往往存在特征条件多等问题,难以应用于存在强交互的密集交通流场景.基于此,对隧道入口车辆换道特性进行分析,并提出一种车辆换道行为预测的模型优选超参数优化方法.首先,基于车辆轨迹数据,从换道位置和持续时间角度进行小轿车和货车换道特性分析;然后,遴选含3 570个样本的轨迹数据,进行描述性统计,并对车辆换道行为进行分类,标记出向左和向右换道样本2 265、1 305组;最后,针对经典逻辑回归和机器学习方法局部最小化以及过拟合问题,优选出Adaboost换道预测模型,采用网格搜索(Grid Search, GS)超参数优化方法实现超参数自动寻优,提出一种GS-AdaBoost优化模型的车辆换道意图预测方法.结果显示:与经典DT(Decision Tree, DT)换道预测模型相比,GS-AdaBoost优化在模型精确率(Precision)、召回率(Recall)、准确率(Accuracy)、F1值(F1score)、特异度(Specificity)方面分别提升了6.18%、17.83%、12.97%、29.38%、8.88%,优选模型性能得到有效提升,表明GS-AdaBoost集成学习优化模型具有较好的鲁棒性和适用性,可为车辆换道意图预测多分类问题提供一种理论方法.Accurately estimating the behavior of vehicles at tunnel entrances is of great significance for improving tunnel traffic safety and efficiency.Existing vehicle behavior prediction methods often overlook the interaction information between vehicles,and the vehicle motion data typically suffer from issues such as high feature dimensionality,making it challenging to apply them to densely populated traffic scenarios with strong interactions.To address this,this study analyzes the lane-changing characteristics of vehicles at tunnel entrances and proposes a model parameter optimization method for predicting lane-changing behavior.Firstly,based on vehicle trajectory data,the lane-changing characteristics of passenger cars and trucks are analyzed from the perspectives of lane-changing position and duration.Then,a subset of trajectory data containing 3,570 samples is selected for descriptive statistics,and the lane-changing behaviors of vehicles are classified.A total of 2,265 and 1,305 samples are labeled as left and right lane changes,respectively.Lastly,to overcome the limitations of classic logistic regression and machine learning methods such as local minima and overfitting,an optimized Adaboost lane-change prediction model is selected.The Grid Search(GS)method is employed to automatically optimize the hyperparameters,resulting in the proposed GS-AdaBoost optimized model for predicting vehicle lane-change intentions.The results show that compared to the classic Decision Tree(DT)lane-change prediction model,the GS-AdaBoost optimized model exhibits improvements in precision,recall,accuracy,F1-score,and specificity by 6.18%,17.83%,12.97%,29.38%,and 8.88%,respectively.The optimized model demonstrates enhanced performance,indicating that the GS-AdaBoost ensemble learning optimization model is robust and applicable.It provides a theoretical approach for multi-class lane-change intention prediction in vehicles.

关 键 词：交通工程 隧道入口 换道意图预测 机器学习 网格搜索 ADABOOST 

分 类 号：U458[建筑科学—桥梁与隧道工程] U491[交通运输工程—道路与铁道工程]