| In recent years,traffic accidents happen frequently,which has brought a lot of losses to the society.Traffic accidents caused by tired driving are increasing year by year.For big car drivers,fatigue driving is affecting their personal safety.Therefore,now many scholars have done a lot of research on the detection of fatigue driving.The research on fatigue driving can be roughly divided into three aspects: through the driver’s physiological characteristics monitoring;According to the vehicle motion characteristics monitoring;Monitoring of drivers’ facial features.It is necessary to wear monitoring equipment to judge whether a driver is tired based on physiological characteristics,which will affect the driver’s driving experience.Moreover,the equipment is expensive and easily damaged.Motion feature monitoring is mainly to judge the fatigue state by monitoring the physical information of the vehicle,such as the vehicle’s driving track,swing range,and the hand pressure on the steering wheel.Although this method will not affect the driving experience of the driver,it will be affected by some external factors such as the driver’s driving habits and special road conditions.Under normal circumstances will not be used as the main basis of judgment.Most of the current studies are based on facial information to conduct fatigue detection.However,these methods all have errors in feature extraction,leading to problems such as low efficiency of feature extraction,low detection accuracy and poor model robustness.And in the subsequent fatigue judgment,the evaluation parameters are single.In view of this,this paper proposes a fatigue driving detection algorithm based on neural network and carries out simulation verification.The innovation points of this paper are as follows:(1)An optimization algorithm of face target detection based on YOLOV3-Tiny is proposed.Combined with Mobile Netv3 network,the YOLOV3-Tiny target detection model is improved as the driver face detection architecture.Then,Wider Face(Face Detection Data Set and Benchmark),an open source Data Set,is used to train the FACE target Detection model.Compared with the existing target detection algorithms of the same kind,the algorithm proposed in this paper has higher recognition accuracy,simplified network structure,less computation,and convenient model transplantation to mobile devices such as mobile phones.(2)A face motion feature extraction algorithm based on eye feature vector(EFV)and mouth feature vector(MFV)is proposed.For the extraction of facial motion features,this paper firstly carries out face feature localization and feature vector extraction based on the dlib toolkit,and then sets two parameters,EFV and MFV,to carry out parameter extraction of the features located by the dlib toolkit,for evaluating the driver’s eye state and mouth state.(3)A fatigue evaluation model based on SVM algorithm is proposed.Most of the existing driver fatigue evaluation algorithms are based on PERCLOS.This algorithm is based on the driver’s eye conditions to determine the fatigue.However,if the driver’s eyes are too small,there will be a higher rate of misjudgment.Similarly,algorithms based on yawning frequency are related to the size of the driver’s mouth.Therefore,in view of this misjudgment phenomenon,this paper designed a SVM classifier algorithm for the classification of drivers’ eyes and mouths.It can determine whether the driver is tired according to the size of the driver’s eyes and mouth at the beginning,which effectively reduces the probability of misjudgment.(4)Driver identity information database establishment.Existing machine learning algorithms considering individual characteristics usually train classifiers by initialization before system startup,and they need to be reinitialized every time the driver is changed.Such algorithms are not only inefficient,but also unable to ensure that every initialization works well.Therefore,this paper constructs the driver identity information database,which includes three types of driver identity information: driver biometric identification classifier,driver eye status classifier and driver mouth status classifier.The algorithm trains the classifier in advance and stores it in the driver identity database.The identity is then verified by calling the driver classifier before the system runs.This not only simplifies initialization,but also avoids inaccuracies caused by manually entering the identity. |
