基于改进卷积神经网络的驾驶员眼睛状态识别  

Driver’s Eye State Recognition Based on Improved Convolutional Neural Network

作　　者：陈仁祥 胡超超 胡小林 赵树恩[1] 蔡东吟 CHEN Renxiang;HU Chaochao;HU Xiaolin;ZHAO Shuen;CAI Dongyin(Chongqing Engineering Laboratory for Transportation Engineering Application Robot,Chongqing Jiaotong University,Chongqing 400074,China;Chongqing Innovation Center of Industrial Big-Data Co.,Ltd.,Chongqing 400056,China)

机构地区：[1]重庆交通大学交通工程应用机器人重庆市工程实验室,重庆400074 [2]重庆工业大数据创新中心有限公司,重庆400056

出　　处：《铁道学报》2023年第6期50-57,共8页Journal of the China Railway Society

基　　金：国家自然科学基金(51975079);重庆市教委科学技术研究项目(KJZD-M202200701);重庆市研究生联合培养基地(JDLHPYJD2021007);重庆市专业学位研究生教学案例库(JDALK2022007)。

摘　　要：为准确、高效识别驾驶员眼睛状态,提出一种基于改进卷积神经网络(Improved Convolutional Neural Networks,ICNN)的驾驶员眼睛状态识别方法。在LeNet-5网络的基础上采用多个小卷积层堆叠替换一个大卷积层的策略,减少参数量和浮点运算数的同时增强网络对眼睛图像的特征提取能力;在卷积层和池化层之间嵌入高效通道注意力(Efficient Channel Attention,ECA)模块,使网络突出眼睛图像中重要通道特征并弱化非重要通道特征,完成ICNN的构建;利用ICNN准确、高效自学习图像中有效眼睛状态特征信息的特点,实现端到端的驾驶员眼睛状态识别。通过在两个公开和一个实测的眼睛数据集上进行对比实验,验证卷积层堆叠替换和嵌入ECA模块的有效性,所提方法具有更高的训练效率和眼睛状态识别准确率。In order to recognize the driver’s eye state accurately and efficiently,an eye state recognition method based on improved convolutional neural networks(ICNN)was proposed.Firstly,on the basis of LeNet-5 network,the strategy of replacing one large convolutional layer with multiple small convolutional layers was adopted to reduce the number of parameters and floating point operations while enhancing the network′s ability to extract features of eye images.Secondly,the efficient channel attention(ECA)module was embedded between the convolutional layer and the pooling layer to highlight the important channel features and weaken the non-important channel features in the eye images,before completing the construction of the ICNN.Finally,taking the advantage of ICNN’s feature of accurately and efficiently self-learning the effective eye state characteristic information in eye images,the end-to-end driver’s eye state recognition was realized.Through comparative experiments on two public and one measured eye data sets,the effectiveness of convolutional layer stack replacement and embedding of ECA modules was verified,proving higher training efficiency and eye state recognition accuracy of the proposed method.

关 键 词：驾驶疲劳 眼睛状态识别 卷积神经网络 高效通道注意力 

分 类 号：V448.151[航空宇航科学与技术—飞行器设计]