Research On Fatigue Driving State Recognition Based On Neural Network

Fatigue driving will significantly reduce the driver's vigilance,increase reaction time,and seriously endanger road safety.It is now listed as the three major causes of traffic accidents with overload and speed.With a focus on the perennial occurrence of traffic crashes brought about by fatigue driving,the convolutional neural network is applied to the state recognition of fatigue driving in this paper.The main research contents are as follows:First,the multitask cascaded convolutional networks(MTCNN)architecture is used to complete face detection and face alignment at the same time,its performance is improved by rough detection and fine extraction.To further improve the detection accuracy,avoid the influence of driver's posture changes and lighting and occlusion in unconstrained environment,MTCNN is improved and the parameters of O-Net is optimized.The experimental results show that the accuracy increases from 97.26% to98.83%,with an increase of 1.57 percentage points,under the premise that the loss of feature points only increases a little.Secondly,Eye and Mouth-CNN(EM-CNN),is proposed to detect the states of the binocular and mouth.In order to reduce the influence of factors such as changes in lighting,differences in sitting posture and occlusion of glasses,and improve the adaptability to complex environments,a total of 4000 images of a real driving environments is collected.Experimental results demonstrate that the presented EM-CNN can efficiently distinguish driver fatigue status using driving images,which outperforms other CNN-based methods i.e.,AlexNet,VGG-16,GoogLeNet,and ResNet50,showing accuracy,sensitivity and specificity rates of 93.62%,93.64%and60.88%,respectively.When EM-CNN adopts Adam optimization algorithm,the accuracy,precision,recall and F-scores are all above 94%.The average AUC value of EM-CNN for mouth state classification is 99.8%,which is better than binocular.Finally,a fatigue driving judgment model based on the fusion of Percentage of Eyelid Open over Pupil over Time(PERCLOS)and Percentage of Mouth Closure over the Pupil over Time(POM).13 videos of disparate drivers in a real driving environment is collected and converted into frame images for identification.Experimental results indicate that when PERCLOS reaches 0.25 and POM reaches 0.5,the driver can be considered to be in a fatigue state.