基于深度学习的疲劳驾驶检测

AbstractThe function of the eyes is not only a window to see the world,but also acarrier of expression.Through the eyes,we can show a variety of States,and theeyes also have the function of conveying feelings,reflecting people's mentalstate and so on.For example,when people are tired,they will blink frequently.When people are full of spirit,their eyes will be more vivid.The frequency ofblinking can reflect whether people are in the state of fatigue to a certain extent.This paper is based on the frequency of opening and closing eyes to determinewhether the driver is driving fatigue.In the state of driving the vehicle,we canexclude the influence of body factors on the blink frequency,so this paper usesthe method based on deep learning to detect the opening and closing frequencyof eyes to detect fatigue driving.The main work of this paper is as follows:(1)The model modification,model training,face region detection and eyeregion location of multitask convolutional neural network (mtcnn)based ondeep learning.MTCNN is chosen here because it has the advantages of highrecognition rate and fast recognition speed.It has three layers of cascadedconvolutional neural network PNET,rnet and onet,which optimizes and selectsfacial features from the input data layer by layer,which is essentially featureclassification and feature regression.Finally,the features of the face area and thefeatures of the eyes,nose and mouth can be obtained,and then through nonmaximum suppression (NMS),the facial features can be obtained.In this paper,the mtcnn model is modified to get the offset of the upper left and lower rightfeature points of the eye region from the output of the new cascaded networklayer,which is the landmark regress.(2)After the mtcnn mentioned above is improved to obtain the facial featurearea and eye feature area,the image of eye feature area is cropped and input intothe CNN layer of fatigue judgment to determine whether the person has fatiguesymptoms.This layer model mainly plays the role of classification to solve the