基于一维卷积神经网络的驾驶人身份识别方法  

Driver Identification Based on 1-D Convolutional Neural Networks

作　　者：胡宏宇[1] 刘家瑞 高菲[1] 高振海[1] 梅兴泰 杨光[1] HU Hong-yu;LIU Jia-rui;GAO Fei;GAO Zhen-hai;MEI Xing-tai;YANG Guang(State Key Laboratory of Automotive Simulation and Control,Jilin University,Changchun 130022,Jilin,China;School of Control and Computer Engineering,North China Electric Power University,Baoding 071003,Hebei,China;GACR&D Center,Guangzhou Automobile Group Co.Ltd.,Guangzhou 510000,Guangdong,China)

机构地区：[1]吉林大学汽车仿真与控制国家重点实验室,吉林长春130022 [2]华北电力大学控制与计算机工程学院,河北保定071003 [3]广州汽车集团股份有限公司广汽研究院,广东广州510000

出　　处：《中国公路学报》2020年第8期195-203,共9页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(51675224,51775236,U1564214);国家重点研发计划项目(2018YFB0105205,2017YFB0102600)。

摘　　要：近年来,智能网联汽车(ICV)已成为智能工业时代最有前景的发展方向。作为现代移动的重要模式,ICV的设计和开发越来越强调个性化需求。提出一种仅使用车载CAN总线行车状态数据,基于深度学习的驾驶人身份识别通用框架。首先采集20名驾驶人在固定试验路线下,包括不同道路类型、不同交通条件下的自然驾驶行车状态数据集;其次对9种类型的CAN信号行车数据进行数据清洗与重采样,构建数据样本集。搭建了由卷积层、池化层、全连接层、SoftMax层构成的一维卷积神经网络(1-D CNN)驾驶人身份识别模型,并且使用Adam算法、L2正则化、Dropout、小批量梯度下降等方法对模型性能进行优化。算法验证过程中,探讨了模型卷积核占比、卷积核数量、卷积层层数、全连接层节点规模对模型识别准确率的影响,进而对模型结构参数进行优选。进一步地,将该算法与K近邻(KNN)、支持向量机(SVM)、多层感知器(MLP)等传统机器学习方法及深度学习算法长短时记忆网络(LSTM)进行对比分析,同时探究样本时间窗口大小、样本数据重叠度、驾驶人数量对模型识别结果的影响。在数据时间窗口为1s、数据重合度80%的条件下,对20名驾驶人进行识别,评价指标宏观F1分数可达99.1%,表明该模型表现明显优于其他对比模型算法,其对驾驶人身份识别表现稳定,鲁棒性强。Currently,the intelligent connected vehicle(ICV)is the most promising direction in the era of intelligent industry.As an important mode of modern mobility,ICVs increasingly emphasize individualized needs in the process of design and development.This paper proposes a general deep-learning framework for driver identification,relying on onboard CAN-bus driving data.First,naturalistic driving data of 20 drivers were collected under a fixed testing route with different road types and different traffic conditions.Second,data resampling was performed on nine types of CAN signal driving data to construct the data sample set.A one-dimensional convolutional neural network driver-identification framework was built,consisting of the convolutional layers,pooling layers,a fully connected layer,and a SoftMax layer.The Adam algorithm,L2 regularization,Dropout,and mini-batch gradient descent were utilized to improve the accuracy and increase the speed of the training process.The convolution kernel ratios,number of convolution kernels,number of convolution layers,and the size of the fully connected layer nodes were optimized by discussing their influences on the model accuracy.Further,the algorithm was compared with traditional machine-learning methods such as K-nearest neighbor,support vector machine,multi-layer perceptron,and another deep-learning algorithm long shortterm memory.The effects of sample time window size,sample data overlap,and number of drivers on the model identification results were discussed.For a sample time window size of 1 s and an overlap of 80%,the evaluation index Macro F1 score reaches 99.1% with the identification of 20 drivers.The results show that the model performance is significantly better than other comparative model algorithms,and it is stable and robust for driver identification.

关 键 词：汽车工程 智能网联汽车 一维卷积神经网络 驾驶人身份识别 行车数据 深度学习 

分 类 号：U461.91[机械工程—车辆工程]