Research On Driver Fatigue Monitoring Algorithm Based On Deep Learning

With the development of the social economy and the automobile industry,automobiles have gradually transformed into ordinary consumer goods and entered people’s daily lives,which brings convenience to people’s daily travel,but at the same time makes road traffic accidents happen more frequent.Studies have shown that a considerable part of traffic accidents are caused by driver fatigue and inattention.Therefore,it is important to develop a system that can monitor the driver’s status in real time and issue early warnings in time of abnormality.The vision-based driver status monitoring system has been widely studied by many scholars du e to its simple equipment and high detection accuracy.In recent years,with the hot developmen t of deep learning in the image processing field,the detection performance of such met hods has been improved to another level.This paper aims to solve the prob lems and deficiencies existing in the previous vision-based driver fatigue monitoring algorithms by using deep learning-based algorithms,and propose a driver fatigue monitoring system with high accuracy and robustness.The specific research contents of this paper are as follows:1)This paper designs a face detection network based on the SSD algorithm for driver’s face detection task.The detection accuracy of the network exceeds the stateof-the-art methods.It can run at a speed of 60 FPS on the edge computing device NVIDIA Jetson Nano and the model size is only 2.7MB.The network can classify faces according to the yaw angle of the face,which can not only assist in judging the driver’s distraction,but also solve the problem that the accura cy of feature extraction decreases under large yaw angles.2)This paper designs a multi-task feature extraction network based on the adversarial autoencoder to perform facial landmarks detection and head pose estimation at the same time.On the 300 W test set,the average errors of the facial landmarks detection task and the head pose estimation task are respectively 4.14(NME)and 3.27°,the accuracy of each task reaches the level of state-of-the-art methods and the network runs at a speed of 166 FPS on the GTX1070 platform.The network has the characteristics of few-shot learning,and only needs 10 training images to achieve the facial landmarks detection accuracy of advanced machine learning methods.The network has the characteristics of face generation,which can generate faces and label them uniformly,which can be used for data supplementation.3)In terms of driver state determination algorithm,this paper designs and calculates a total of 20 characteristic parameters of eyes,mouth and head by using the head pose and face landmark information obtained by the feature extraction algorithm combined with the classification results of face detection network.Those characteristic parameters are set as input to explore the effect of artificially setting judgment rules(threshold method),SVM classifier,fully connected network model and long-term memory network model in fatigue status judgment.Experiments show that the Long Short Term Memory network has the best result on the driver’s status determination.