基于迁移学习和AlexNet的驾驶员行为状态识别方法  

Driver Behavior Recognition Method Based on Migration Learning and AlexNet

作　　者：戎辉[1,2] 华一丁 张小俊[2] 龚进峰[1] 唐风敏[1] 郭蓬[1] 何佳[1] RONG Hui;HUA Yi-ding;ZHANG Xiao-jun;GONG Jin-feng;TANG Feng-min;GUO Peng;HE Jia(China Automotive Technology Research Center Co.,Ltd.1,Tianjin 300300,China;School of Mechanical Engineering,Hebei University of Technology,Tianjin 300401,China;China Automotive Research(Tianjin)Automotive Engineering Research Institute Co.,Ltd.,Tianjin 300300,China)

机构地区：[1]中国汽车技术研究中心有限公司,天津300300 [2]河北工业大学机械工程学院,天津300401 [3]中汽研(天津)汽车工程研究院有限公司,天津300300

出　　处：《科学技术与工程》2019年第28期208-216,共9页Science Technology and Engineering

基　　金：国家重点研发计划(2017YFB0102500);天津市科委人工智能重大专项(17ZRXGGX00130);天津市科委新一代人工智能科技重大专项(18ZXZNGX00230);中国汽车技术研究中心有限公司重点课题(16190125,10318-01)资助

摘　　要：为了解决传统基于神经网络算法的驾驶员行为状态识别系统精度过于依赖大量训练样本的问题,提出将迁移学习理论和Alex Net引入到驾驶员行为状态的识别研究中。首先对驾驶员行为特征及状态进行深入分析,对驾驶员7种驾驶状态进行了定义,构建了驾驶员状态信息采集系统;然后对基于卷积神经网络的驾驶员状态识别方法研究,建立了驾驶员状态数据集,构建了基于Alex Net卷积神经网络的状态监测系统,通过迁移学习完成了卷积神经网络识别模型。最后通过实验验证了提出的驾驶员状态识别算法对7种驾驶员状态识别的有效性。实验表明:该系统准确率达到97. 8%,且在实验设备中运行速度达到70帧/min,满足较高的准确率要求与实时性要求。In order to solve the problem that the accuracy of the traditional neural network-based driver behavior state recognition system is too dependent on a large number of training samples,migration learning theory and AlexNet were adapted to research on into the identification research of driver behavior state. Firstly,the driver’s behavior characteristics and state were analyzed in depth,the driver’s seven driving states were defined,the driver state information acquisition system was constructed. Then,the driver state recognition method based on convolutional neural network was studied to establish driving. Based on the state data set,a condition monitoring system based on AlexNet convolutional neural network was constructed. The convolutional neural network recognition model was completed through migration learning. Finally,the effectiveness of the driver state recognition algorithm proposed was verified for seven kinds of driver state recognition. Experiment results show that the accuracy of the system reaches 97. 8%,and the running speed in the experimental equipment reaches 70 frames/min,which meets the high accuracy requirements and real-time requirements.

关 键 词：驾驶员状态 迁移学习 AlexNet 卷积神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]