Research On Fatigue Detection Algorithms Based On Driver's Facial Fusion Features

In recent years,with the development of society,the rapid expansion of road transportation networks and the popularity of private cars,road trips have become an indispensable part of people's daily life,followed by urban traffic safety issues.And every major traffic accident is accompanied by immeasurable loss of life and property.Statistics show that fatigue driving is one of the important causes of traffic safety accidents.Therefore,it is of great practical significance to research and develop a set of efficient,universal and accurate fatigue driving detection systems to reduce the incidence of traffic accidents caused by fatigue driving.At present,there are many fatigue testing equipments on the market.After research and analysis,most of these devices have the following problems: the accuracy rate is greatly affected by illumination factors.the facial positioning and tracking algorithm is less efficient and less robust,what cause the system's performance leave much to be desired.the simplified fatigue judgment mechanism,insufficient fatigue information,and the multisource feature is required to jointly characterize the driver fatigue state.This paper realizes the detection of fatigue driving behavior based on the fatigue characteristics of driver's face by means of the excellent representation ability of in-depth learning in the field of image.The specific work of this paper includes: illumination evaluation and correction of input image,detection and tracking of driver's face position,facial key position and fatigue feature extraction,feature fusion and fatigue determination,and simulation experiment of the system based on wxPython,a open source library.Firstly,this paper proposes an illumination evaluation algorithm to evaluate the illumination of the frame image of the input video stream,and to screen out images with uneven illumination and unsatisfactory illumination intensity.By studying the existing classical illumination compensation algorithm,and based on this,an image illumination correction algorithm based on dynamic parameters is proposed to correct the selected frame images with poor illumination conditions.The corrected image has been greatly improved.Secondly,this paper improves the face detection model MTCNN according to the application scenario of fatigue driving detection.Reduce the computational complexity of the model and improve the detection efficiency by reducing the dimension of the channel.By studying the law of the driver's facial imaging in the cockpit,the input image pyramid of the MTCNN is compressed,which reduces the input data volume of the model,improves the efficiency of the model,and eliminates the interference caused by the rear passengers to the driver's face detection task.Then,for the face detection task in the continuous video stream frame image,we research and analyze the related tracking algorithm,and propose the face tracking algorithm based on MTCNN,the face tracking algorithm based on O-Net and the KCF tracking algorithm based on CNN feature.The experimental verification results show that the efficiency is improved compared with the traditional tracking algorithm.The introduction of the tracking algorithm makes the time-consuming task of the face positioning task greatly compressed,which greatly improves the efficiency of the system.Then,to realize the rapid extraction of driver's facial fatigue characteristics.based on CNN,the face key model was proposed.The trained CNN model can quickly and accurately detect the position of the key points from the input face image.The experiments show that the key point error of the model in the 96*96 face image is about 1 pixel.Based on the face key model,the segmentation of the face-related organs and the extraction of the corresponding fatigue features can be quickly realized,mainly the fatigue characteristics of the eyes,head and mouth.Finally,based on the error Back Propagation neural network,this paper designs the network structure and uses the fatigue feature data set extracted from the experimental samples to complete the model training.Through the network model to fuse the fatigue characteristics,the driver's fatigue coefficient is output,and the cross-validation results show that the accuracy of the model is as high as 98.81%.In this paper,the algorithm introduced in each chapter is implemented in Python language,and the algorithm model of each stage is integrated.The graphical interface of the fatigue driving detection system is built by wxPython.The system can calculate the relevant fatigue feature values in real time,locate and track the driver's face position,and detect the driver's face key point coordinates.The test results show that the driver face tracking loss rate under integrated illumination is not higher than 3.4912%,which can provide reliable fatigue characteristics to the system.The comprehensive processing rate is above 10 frames/second,real-time performance is satisfactory.The average accuracy rate can reach 95.71%,and the average accuracy under ideal illumination conditions reaches 97.47%,which can accurately detect the fatigue state of the driver.The accuracy and efficiency of the related algorithms have also been verified.