Research On Face Fatigue Information Detection Based On Deep Learning

With the continuous progress of science and the sustainable development of economic society,the vehicle population in our country is growing rapidly,then the traffic accidents are also increasing.Statistics show that fatigue driving is one of the important reasons caused the traffic accidents.If we could detect the driver’s fatigue state in time and warn the driver,the traffic accidents caused by fatigue driving would be reduced.Currently,the research field of driver fatigue detection,contact fatigue detection obtains the physiological indicators of driver,such as pulse and electroencephalogram during the driving state for fatigue detection,which has the highest accuracy.However,drivers need wear relevant medical detection equipment when they are driving,which brings high economic cost for driver and affects the normal driving state.Non-contact fatigue detection is more friendly.This method is based on vehicle maneuvering behavior detection,because the driver has different maneuvering behaviors,so the accuracy of detection is not high.As an assistant driving system,vehicle running status detection has been widely used at home and abroad in business.Fatigue detection is based on facial expression recognition,a computer vision detects driver’s facial information,which is high accuracy,simple equipment and low cost.This article studies the direction of fatigue state determination which is based on facial fatigue information,the main contents as follows:(1)Face detection.This article uses MTCNN net work for face detection in order to overcome the adverse factors influence face detection,such as driver posture change,complex light conditions and face occlusion.In order to solve that the real-time fatigue detection of vehicle detection equipment is weak,improved and optimized the MTCNN network structure.Firstly,the lightweight network Mobile Net is introduced with a step size of 3 × 3 which is a deep detachable convolution layer to replaces the original pooling layer in R-Net and O-Net of MTCNN subnetworks.Secondly,optimized the parameters of two subnetworks.Choose the face datasets of WIDER FACE and Celeb A to train and validate the improved MTCNN network.The results show that the accuracy is improved from 97.38% to 98.16% of the improved MTCN N network,and the loss value is only increased by 0.1%.The average detection speed of the improved MTCNN network is increased by 20 MS compared with the original one during the self-made video dataset experiment.The improved MTCNN network is superior to the original network in detection accuracy and speed.(2)Fatigue feature extraction.On the basis of face detection,this paper selects3000 samples from Celeb A dataset to train the cascade regression algorithm ERT,completes high-precision face tagging for 68 key points,and provides accurate feature point location for subsequent feature extraction.To deal with the problem that the low accuracy of single feature fatigue detection,the pepper is based on the P80 criterion of PERCLOS algorithm,EAR is calculated using 12 feature point coordinates of eyes.Blink detection is achieved by EAR threshold,and blink frequency is calculated by counting the number of blinks per unit time.MAR was calculated using eight feature points of mouth contour.Yawning was detected using MAR value and double threshold of mouth opening time.Yawning frequency was also calculated by counting yawning times in unit time.Fourteen 2D facial key points are selected to map to three-dimensional space,matched the 3D face model,solved the coordinate system transformation relationship.The head posture angle is obtained from the rotation matrix to get multi-fatigue features of eyes,mouth and head.(3)Multi-characteristic fatigue evaluation.Collect seven fatigue feature parameters,including eyelid closure,blink frequency,mouth opening,yawn frequency and head pitch,yaw,roll.In order to solve the problem of inaccurate determination of fatigue when multiple fatigue features are judged,the independency of each Fatigue Feature and its contribution to fatigue determination are improved.This paper uses D-S evidence theory to discriminate fatigue driving status.Two D-S models are designed according to different feature grouping modes.All feature groupings in the model are trained by SVM to obtain a posterior probability.Evidence is fused according to Dempster rules to output the final decision results.Th ese two models were trained and validated using Yaw DD dataset,and the accuracy of the model grouped according to the evidence by organs was up to 96.85%.The experimental results show that the fatigue detection algorithm is based on multi-feature fusion has a high accuracy and is superior to the detection method which is based on single Fatigue Feature and the detection method which is based on simple multi-feature fusion.Finally,the algorithm in this paper is integrated and visualized using Python language then Tensor Flow framework to complete the fatigue driving detection system.