Research On Dangerous Driving Behavior Detection Technology Based On Deep Learning

In recent years,the rapid development of the transportation industry has facilitated people’s way of travel,but the incidence of traffic accidents has been increasing.According to a large number of research reports,the major causes of traffic accidents are fatigue driving and distracted driving.To solve these two dangerous driving detection problems,researchers at home and abroad have proposed detection methods based on drivers’ physiological features,facial features and vehicle driving characteristics.The detection method based on the driver’s physiological characteristics and vehicle driving characteristics often affects the driver’s driving condition,with poor comfort and low accuracy.Although the detection accuracy is improved,the detection speed is slow and the model memory is large.In order to solve the above problems,this thesis proposes a dangerous driving detection algorithm based on deep learning,which has high accuracy and real-time performance.The main work of this thesis is as follows:(1)Driver’s face detection and key point detection and extraction.In the face detection task,in order to solve the problem of slow model detection speed,this thesis proposes a lightweight network based on MTCNN,which replaces the ordinary convolution in P-Net,R-Net and O-Net networks with deeply separated convolution,and adds the Dropout layer in front of the fully connected layer of the O-Net network to prevent overfitting phenomenon.The detection speed and accuracy of the proposed algorithm are verified by experiments.In the task of face key point detection and extraction,in order to solve the problems of the original model’s high attention to exaggerated samples and slow convergence of loss function,this thesis improves the auxiliary network and loss function based on the PFLD algorithm.First,the loss function of face key point is modified to the Wing loss function,which can well solve the problem of "outliers".Then,the head pose estimation of the original auxiliary network was replaced by the position estimation of the eyes and mouth.The positions of the eyes and mouth were obtained by averaging the positions of the key points of face 68 and the positions estimated by the auxiliary network,and the corresponding regions were extracted according to the key points of the eyes and mouth.(2)Recognition and judgment of fatigue characteristic state.In the task of fatigue characteristic state recognition,in order to improve the accuracy and reasoning speed of the model,three different attention mechanisms were added based on the YOLOv5 s algorithm in this thesis.Through experiments,the influences of different mechanisms on the size,accuracy and detection speed of the model were compared.Finally,a network with good performance was selected as the eye and mouth state recognition network.Euler Angle is used to determine the head attitude.In the task of fatigue state determination,a single facial feature is no longer used,and a new fusion method is proposed in this thesis.The weighted average sum of blink frequency,PERCLOS parameter,maximum eye closing time,yawning frequency and nodding frequency parameters is carried out to determine the fatigue level of drivers according to the calculated fatigue values.(3)Distracted driving detection.In this thesis,ten driving behaviors are selected:safe driving,talking on the phone with both hands,texting with both hands,drinking water,adjusting vehicle equipment,turning the head back,arranging makeup,and communicating with passengers.A Resnet50-CA network was proposed by adding the attention mechanism with better selection performance to the Resnet50 network.The model detection effect before and after the improvement was visualized by Grad-CAM thermal map.