不同行为模型框架下的数据驱动类人驾驶模型比较  

Comparison of Data-Driven Human-Like Driving Models Under Different Behavior Model Frameworks

作　　者：黄玲[1,2,3] 黄子虚 吴泽荣 游峰 崔躜[1] 曾译萱 钟浩川 HUANG Ling;HUANG Zixu;WU Zerong;YOU Feng;CUI Zhuan;ZENG Yixuan;ZHONG Haochuan(School of Civil Engineering and Transportation,South China University of Technology,Guangzhou 510640,Guangdong,China;Jiangsu Key Laboratory of Urban ITS,Southeast University,Nanjing 210096,Jiangsu,China;State Key Laboratory of Subtropical Building Science,South China University of Technology,Guangzhou 510640,Guangdong,China)

机构地区：[1]华南理工大学土木与交通学院,广东广州510640 [2]东南大学现代城市交通技术江苏高校协同创新中心,江苏南京210096 [3]华南理工大学亚热带建筑科学国家重点实验室,广东广州510640

出　　处：《华南理工大学学报（自然科学版）》2022年第10期1-10,共10页Journal of South China University of Technology(Natural Science Edition)

基　　金：广东省区域联合基金资助项目(2020B1515120095);广州市重点领域研发计划项目(202007050004);广东省普通高校特色创新类项目(2019KTSCX007);华南理工大学亚热带建筑科学国家重点实验室开放研究项目(2020ZB20)。

摘　　要：传统的驾驶行为模型框架把驾驶行为划分为跟驰行为和换道行为两大类,并分别进行模型构建;而综合驾驶行为模型框架则认为跟驰行为和换道行为密不可分,因此将所有驾驶行为看作一个整体来进行建模。文中基于这两种行为模型框架,对数据驱动类人驾驶模型的性能进行分析。首先,建立综合驾驶行为模型框架和跟驰换道组合模型框架,并根据驾驶过程中的影响因素,确定模型的输入和输出。其次,提出了基于跟驰、换道和意图识别模块的两种跟驰换道组合方式:判别组合和概率组合。随后,对原始数据集进行处理筛选,构建综合驾驶行为、跟驰行为、换道行为和意图识别4个样本库,分别用于对相应行为模块进行训练和标定。最后,将两种跟驰换道组合模型与综合驾驶行为模型进行模型精度、安全性、鲁棒性和迁移性比较。结果表明:在模型输入输出、参数标定流程和样本数据库一样的情况下,基于长短期记忆神经网络(LSTM)的类人驾驶模型精度优于基于FNN的模型,其中基于LSTM的模型均方误差可达到0.227 m^(2),基于FNN的模型均方误差为0.470 m^(2)。而在基于LSTM的模型中,采用跟驰换道组合模型框架的模型比采用综合驾驶行为模型框架的模型具有更好的鲁棒性和迁移性,其中跟驰换道组合模型在±10%噪声下的鲁棒性均方误差可达到1.383m^(2),迁移性均方误差可达到0.462 m^(2);综合驾驶行为模型在±10%噪声下的鲁棒性均方误差为2.314m^(2),迁移性均方误差为0.484m^(2)。The traditional driving behavior model framework divides the driving behaviors into car-following and lane-changing,which are modeled separately.While the integrated driving behavior model framework believes that car-following and lane-changing are inseparable,so all driving behaviors are modeled as a whole.Based on these two behavioral model frameworks,this paper analyzed the performance of the data-driven human-like driving models.Firstly,it established integrated driving behavior model framework and car-following lane-changing combined model framework and then determined the input and output of the models according to the influencing factors in driving.Secondly,two combinations of car-following,lane-changing and intention recognition modules were proposed:discriminative combination and probability combination.Subsequently,the processing of the original data were carried out to build integrated driving behavior,car-following,lane-changing,and intention recognition datasets,which were used to train and calibrate the corresponding modules.Finally,the study compared the performance of the two combination models with the integrated driving behavior model in various aspects,including model accuracy,safety,robustness and migration.The results show that,when the model input and output,the parameter calibration process and the dataset are the same,the accuracy of the human-like driving model based on long short-term memory neural network(LSTM)is better than the model based on FNN.The mean square error of the model based on LSTM can reach 0.227 m^(2),and the mean square error of the model based on FNN is 0.470 m^(2).Within the LSTM-based model,the model using the car-following lane-changing combined model framework has better robustness and transferability than the model using the integrated driving behavior model framework.For the carfollowing lane-changing combined model,the mean square error of±10%noise robustness can reach 1.383 m^(2),and the mean square error of transferability can reach 0.462 m^(2).For the integr

关 键 词：数据驱动 类人驾驶 综合驾驶行为 换道行为 跟驰行为 长短期记忆神经网络 

分 类 号：U491[交通运输工程—交通运输规划与管理]