Research On Fatigue Driving Detection Method Of Mobile Edge Terminal

With the continuous improvement of people’s living standards,more and more people choose motor vehicles to travel,resulting in frequent traffic accidents.Among them,fatigue driving is one of the main reasons for causing traffic accidents.In most cases,it is difficult for the driver to realize that he is in a fatigued driving state.Therefore,real-time monitoring of the driver’s fatigue status and timely warning when the driver is in a fatigued state has important meaning of reducing the incidence of traffic accidents and ensuring people’s safe travel.Most of the existing fatigue driving detection methods do not consider the problem of limited computing resources on the mobile edge terminal.In order to independently complete the task of fatigue detection on low-computing in-vehicle embedded devices,this paper proposes a fatigue driving detection of mobile edge terminal.The main research contents are as follows:(1)Research on driver face detection and state classification method of facial feature area.Multi-Task Cascaded Convolutional Networks(MTCNN)has high accuracy in face detection tasks,however,its detection on mobile edge devices with low computing power takes too long.Aiming at this problem,a speed-up method of MTCNN with parameter self-tuning is proposed,which can accurately and efficiently locate the driver’s face area.;and a high-speed Kernel Correlation Filter(KCF)is used to track the feature area,aiming at the problem that KCF continues to track the wrong target after the target is lost,a target self-updating KCF tracking method based on Intersection over Union(IOU)is proposed.Finally,the Res Net18,which has shown superior performance on image classification tasks,is used to classify the state of the driver’s facial feature region to prepare for subsequent fatigue detection.(2)Research on lightweight method of facial feature region state classification model for mobile edge terminal.The facial feature region state classification model based on Res Net18 has a complex structure,a large number of parameters,and a large amount of calculation,its reasoning speed is slow on mobile edge devices with low computing power,and it cannot meet the real-time requirements of fatigue detection tasks.Aiming at the problem,a combined model lightweight method based on Similarity-preserving knowledge distillation and Linearly replaceable filters for deep network channel pruning(LRF)is proposed to lighten the facial feature region state classification model.So that,it ensures high recognition accuracy and has a smaller volume,achieving a balance of inference speed and accuracy.(3)Research on fatigue driving detection method based on facial multi-feature fusion.Aiming at the problem of inaccurate discrimination of a single fatigue feature,the Percentage of Eyelid Closure Over the Pupil over Time(PERCLOS),the Percentage of Mouth Opening Time(PMOT)and the nodding frequency are proposed to characterize a variety of facial fatigue characteristics.The weight of each evaluation index relative to the fatigue state is obtained by the Analytic Hierarchy Process,and a joint evaluation index F,which fuses the three indicators according to the weight,is proposed.Finally,the driver’s fatigue state is detected according to the joint evaluation index F to improve the accuracy and robustness of the fatigue detection method under the premise of real-time detection on mobile edge devices with limited computing and storage resources.