| 实测数据驱动的自动驾驶道路测试驾驶模式辨别方法 |
| |
| 引用本文： | 涂辉招,刘芳丽,崔航,鲍胜,赵瑜,曹寅,曹建永.实测数据驱动的自动驾驶道路测试驾驶模式辨别方法[J].中国公路学报,2021,34(4):231-239. |
| |
| 作者姓名： | 涂辉招 刘芳丽 崔航 鲍胜 赵瑜 曹寅 曹建永 |
| |
| 作者单位： | 1. 同济大学交通运输工程学院, 上海 201804;2. 上海市交通委员会科技信息中心, 上海 200080;3. 上海机动车检测认证技术研究中心有限公司, 上海 201805 |
| |
| 基金项目： | 国家重点研发计划项目（2019YFE0108300）；山西省交控集团重点研究计划项目（20-JKKJ-1） |
| |
| 摘 要： | 自动驾驶道路测试中车企驾驶模式数据具有一定保密性，导致自动驾驶能力难以被客观评估。为此，提出了实测数据驱动的自动驾驶道路测试驾驶模式辨别方法。首先选取数据特征值构建K近邻估计、支持向量机、决策树、随机森林和BP神经网络5种机器学习监督分类模型；其次通过非参数秩和显著性检验确定驾驶模式持续时长阈值，持续时长大于阈值的数据段记录为准确的驾驶模式数据，小于等于阈值的数据段则为驾驶模式待分类数据集；随机选取70%记录准确的驾驶模式数据作为监督分类模型训练数据集，剩余30%作为测试数据集；最后利用正确率、精确率和召回率3个指标评价5种监督分类模型，并选取表现最佳的分类模型用于待分类数据的驾驶模式辨别。基于上海市城市道路和快速路2个道路测试场景共约43.6万条数据，验证驾驶模式辨别方法的有效性。结果表明：随机森林监督分类模型辨别道路测试驾驶模式的效果最佳；城市道路场景和快速路场景待分类数据驾驶模式记录有误率分别达到42.3%和39.4%。实测数据驱动的驾驶模式的辨别与修复，可显著提升评估自动驾驶道路测试驾驶能力的准确度。
|
| 关 键 词： | 交通工程 驾驶模式辨别 机器学习 监督分类 自动驾驶 道路测试 |
| 收稿时间： | 2020-02-23 |
Empirical Data-driven Identification of Driving Modes in Autonomous Vehicle Road Testing |
| |
| TU Hui-zhao,LIU Fang-li,CUI Hang,BAO Sheng,ZHAO Yu,CAO Yin,CAO Jian-yong.Empirical Data-driven Identification of Driving Modes in Autonomous Vehicle Road Testing[J].China Journal of Highway and Transport,2021,34(4):231-239. |
| |
| Authors: | TU Hui-zhao LIU Fang-li CUI Hang BAO Sheng ZHAO Yu CAO Yin CAO Jian-yong |
| |
| Affiliation: | 1. College of Transportation Engineering, Tongji University, Shanghai 201804, China;2. Science and Technology Information Center of Shanghai Municipal Transportation Commission, Shanghai 200080, China;3. Shanghai Motor Vehicle Inspection and Certification Technology Research Center Co. Ltd., Shanghai 201805, China |
| |
| Abstract: | Owing to the confidentiality of the driving mode data of automobile enterprises, objectively evaluating the driving performance of autonomous vehicle in road testing (AVRT) is difficult. Therefore, this study proposes an empirical data-driven approach to identify the driving modes of AVRT. First, the data features are selected to establish five supervised machine learning classification models:K-nearest neighbor estimation (KNN), support vector machine (SVM), decision tree (DT), random forest (RF), and back propagation (BP) neural network algorithms. Second, the threshold values of driving mode durations are determined through non-parametric rank and significance tests. When the duration was greater than the threshold value, the data was considered to be recorded clearly on the driving modes. If the duration is less than, or equal to, the threshold value, the data was labeled as the unclassified dataset, for which the driving modes need further identifications. Seventy percent of the clearly recorded driving mode data were randomly selected as the supervised classification model training dataset, and the remaining 30% were treated as the test dataset. Finally, the five types of supervised classification models were evaluated using three indicators:correctness, accuracy, and recall. The classification model with the best performance was selected to identify the driving mode of the unclassified dataset. The proposed approach on driving mode identification was verified based on 0.436 million road testing records from urban-road and highway scenarios in the city of Shanghai. The results show that RF supervision model is the most suited for driving mode identification in AVRT. The inaccuracy rates for driving modes in the AVRT data for the urban roads and highways were found to be 42.3% and 39.4%, respectively. The driving ability assessment accuracy rate in AVRT is significantly improved by identifying and repairing the incorrectly recorded driving modes in the empirical data. |
| |
| Keywords: | traffic engineering driving modes identifying machine learning supervision classification autonomous vehicle road testing |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《中国公路学报》浏览原始摘要信息 |
|
点击此处可从《中国公路学报》下载全文 |
|
