基于眼睛状态识别的疲劳驾驶检测  

Detection to Fatigue Driving Based on Eye State Recognition

作　　者：徐莲 任小洪 陈闰雪 XU Lian;REN Xiao-hong;CHEN Run-xue(School of Automation and Information Engineering,Sichuan University of Science and Engineering,Yibin 644000,China;Artificial Intelligence Key Laboratory of Sichuan Povince,Zigong 643000,China;School of Computer,Sichuan University of Science and Engineering,Yibin 644000,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,宜宾644000 [2]人工智能四川省重点实验室,自贡643000 [3]四川轻化工大学计算机学院,宜宾644000

出　　处：《科学技术与工程》2020年第20期8292-8299,共8页Science Technology and Engineering

基　　金：四川省教育厅基金(17ZB0302)。

摘　　要：由于光照变化、头部姿态等因素的影响,现有的疲劳驾驶检测精度仍有待提高。针对该问题,提出一种基于迁移学习的眼睛状态识别网络(Gabor and LBP-convolutional neural networks,GL-CNN),该网络是由Gabor特征和LBP特征通过迁移学习加入卷积神经网络(convolutional neural networks, CNN)调制组成的。首先用多任务级联卷积神经网络(multi-task CNN,MTCNN)检测驾驶员的人脸和双眼,然后经过眼睛筛选机制获取待检测的单只眼睛,通过GL-CNN识别眼睛的睁闭状态,最后根据PERCLOSE准则判断驾驶员的疲劳状态。实验结果表明,该算法具有较高的准确率,可以检测多种姿态眼睛的状态,同时满足实时性的要求。Due to the influence of light changes, head posture, and other factors, the accuracy of detection to the eye state in fatigue driving still needs to be improved. Aiming at this problem, an eye state recognition network GL-CNN(Gabor and LBP-convolutional neural networks) was proposed based on transfer learning. The network adds the Gabor and LBP features to the convolutional neural networks(CNN) modulation by transfer learning. First, the state of driver’s face and eyes was detected by the multi-task cascade MTCNN(Multi-Task CNN), and then the single eye was detected and acquired through the eye screening mechanism, and the opening-and-closing state of the driver’s eyes was identified by GL-CNN. Finally, the driver’s fatigue state was determined according to the PERCLOSE criterion. The experimental results show that the algorithm had higher accuracy and can detect the state of eyes in various poses while meeting the real-time requirements.

关 键 词：疲劳驾驶检测 迁移学习 眼睛筛选机制 多任务级联卷积神经网络 眼部状态识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]