基于DCNN的管制员疲劳状态检测  

Controller Fatigue State Detection Based on DCNN

作　　者：梁海军[1] 刘长炎 陈宽明[1] 孔建国[1] LIANG Hai-jun;LIU Chang-yan;CHEN Kuan-ming;KONG Jian-guo(College of Air Traffic Management,Civil Aviation Flight University of China,Guanghan 618300,China)

机构地区：[1]中国民用航空飞行学院空中交通管理学院,广汉618300

出　　处：《科学技术与工程》2021年第35期15277-15283,共7页Science Technology and Engineering

基　　金：中国民用航空飞行学院大学生创新创业训练计划项目(S202010624094);中国民用航空飞行学院科研基金(J2018-60);2020年民航教育培训项目“省级教学实验平台提升建设”(2052036)。

摘　　要：空中交通管制员疲劳工作势必会严重威胁空中交通安全,通过对眼睛睁闭状态判定是现阶段对管制员疲劳检测的一种主要方式。为检测管制员疲劳状态,提出了一种基于迁移学习的DCNN(deep convolutional neural network)眼睛状态识别模型。首先,利用深度级联神经网络的MTCNN(multi-task cascaded convolutional networks)算法检测出管制员面部区域,并实现对面部5个关键点标定和眼睛的定位;然后将检测到的眼睛图像传入到预训练的DCNN眼睛状态分类模型,识别眼睛的睁闭眼状态;最后结合PERCLOS(percentage of eyelid closure over the pupil over time)80指标检测管制员疲劳状态。分别在ZJU、CEW和ATCE数据集上,对DCNN、VGG16、InceptionV3、ResNet504种模型的准确率、损失率和F_(1)分数指标进行对比实验。实验结果表明:在ZJU和CEW数据集上,DCNN眼睛状态分类模型检测准确率为97%,较VGG16、InceptionV3、ResNet50等模型进行眼部状态分类任务,DCNN模型的F_(1)分数有3%~7%的提高。在ATCE数据集上DCNN模型检测准确率达到98.35%,F_(1)分数达到98.06%,验证了DCNN模型的有效性与准确性。The fatigue work of air traffic controllers is bound to seriously threaten the air traffic safety.It is a main way to detect the fatigue of air traffic controllers by judging the state of eyes open and closed.In order to detect the controller's fatigue state,a DCNN(deep convolutional neural network)eye state recognition model based on transfer learning was proposed.Firstly,the MTCNN(multi-task cascaded convolutional networks)algorithm of deep cascaded neural network was used to detect the controller's face area.The five key points of the face are calibrated and the eyes are located.Secondly,the detected eye images are transferred to the pretrained DCNN eye state classification model for recognizing the open and closed eyes.Finally,the controller's fatigue state was detected combined with PERCLOS(percentage of eyelid closure over the pupil over time)80 index.The accuracy,loss rate and F_(1) score of DCNN,VGG16,InceptionV3 and ResNet50 models were compared on ZJU,CEW and ATCE datasets.Experimental results show that on ZJU and CEW datasets,the detection accuracy of DCNN eye state classification model is 97%,compared with VGG16,InceptionV3,ResNet50 models for eye state classification tasks,DCNN model F_(1) score has been improved by 3%to 7%.The accuracy of DCNN model detection reaches 98.35%and the accuracy of F_(1) score reaches 98.06%,which verify the validity and accuracy of DCNN model.

关 键 词：空中交通管制员 卷积神经网络 迁移学习 眼部疲劳检测 

分 类 号：V355.1[航空宇航科学与技术—人机与环境工程] TP391.41[自动化与计算机技术—计算机应用技术]