Study And Application Of Fatigue Detection Of Driver’s Face Video Image Based On Convolutional Neural Network

Driving fatigue is one of the main causes of traffic accidents.It is of important siginificance to detect driver fatigue status in time and alerts properly to fatigue drivers for preventing many accidents,and reducing personal suffering and driving Safely.In the paper,It mainly focuses on the driver fatigue detection in the driver’s face in the complex situation of the driving environment(face occlusion problem),and uses the method of convolutional neural network to detect the fatigue state.This paper mainly studies driver face detection,driving fatigue feature extraction and driving fatigue detection network building.Firstly,the paper deeply analyzed face detection methods and selected MTCNN based on CNN cascade structure as the driver face detection method.Secondly,based on the face detection,utilizing the spatial and temporal components of the video frame data extract the spatial static features and temporal dynamic features of the driver’s face information respectively.Therefore,a two-stream CNN is bulit by improved AlexNet for Driving fatigue detection.The two-stream CNN is a parallel structure of the incorporating of spatial stream network and temporal stream network.Spatial flow network extracts static fatigue features of drivers in current RGB frames.Temporal stream network extracts information of neighboring frames which is represented by optical flow field vector.Features of time stream and spatial stream extraction is fused by convolutional fusion method to detect driving fatigue.Finally,the weighted average method is used to fuse the vehicleroad information to study driving fatigue detection.The face detection experiment compares the detection efficiency of the traditional face detection method(the face detection algorithm based on HOG operator and the Adaboost face detection algorithm based on Haar feature)and the MTCNN method.The experiment show that MTCNN has the best detection effect.The accuracy of face detection using MTCNN achieves 99.7% and the average detection time per frame achieves 12.5 ms.In addition,Driving fatigue experiments show that the accuracy of the two-stream CNN based on AlexNet improvement achieves 92.87%,the accuracy of the single-channel spatial stream network using only static features achieves 87.12%,and the accuracy of the single-channel temporal stream network using only dynamic features achieves 83.5%.The paper have compared and analysed several groups of experiments to modify model parameters,which can ensure the efficiency of the model performance.Finally,the fatigue features of the vehicle-road information were fused to detect driving fatigue,and the detection achieves 94.02%.