Research On Driver Fatigue Detection Method Based On Face Features

As living standards improve and the number of motor vehicles increases year on year,traffic accidents are also on the rise,with driver fatigue accounting for approximately 20% of all accidents.If the driver’s fatigue is determined by technical means and reminded in time,the incidence of traffic accidents can be reduced.The study of driver fatigue detection is therefore of great importance for the protection of people and property.There are three kinds of existing fatigue detection methods: based on physiological characteristics,based on vehicle characteristics,and based on behavioral(facial)characteristics.The method utilized in this thesis for detecting fatigue relies on facial features and employs machine vision.It offers several benefits such as noninvasiveness,practicality,and affordability.Most of the existing fatigue detection methods do not consider real-time performance or are deployed on edge devices.In order to ensure the accuracy and real-time performance of detection,this thesis integrates multiple features to comprehensively judge fatigue,and all designs are as lightweight as possible.The main research contents of this thesis are:(1)Driver’s face landmark detection modelA lightweight dual-channel facial landmark detection model has been developed to address the existing challenges in facial landmark detection models and the need for fatigue detection.The model uses a convolutional neural network architecture,with each channel dedicated to the detection of key points for the eye and mouth regions,respectively.The use of ghost modules in the basic convolutional blocks significantly reduces the number of parameters in the model.The proposed model outperforms most state-of-the-art models,achieving normalized mean errors of 4.06%,5.31%,and 3.92%on the 300 W,WFLW,and FADID datasets,respectively.Due to its training on challenging datasets,the model demonstrates robust performance in detecting facial landmarks under conditions such as varying lighting conditions,unconstrained poses,and occlusions.(2)Driver fatigue state recognitionThis thesis addresses the problem of inaccurate detection in single-feature analysis by integrating multiple sources of information to improve detection accuracy.First,the head pose of the driver is estimated.Then,considering the diversity of human eyes,a Gaussian distribution is fitted to determine the mean and standard deviation for fully open and fully closed eye states.Based on these parameters,a segmented PERCLOS definition is designed.Furthermore,a pose-independent method is proposed to estimate the mouth state,effectively distinguishing between slight yawning and speaking with a smile,which are difficult to differentiate.To overcome the potential real-time limitations of deep learning approaches for fatigue detection,this thesis employs a support vector machine as a classification model for binary classification of driver drowsiness and alertness.The fatigue detection model achieves 88.2% accuracy and 91.7% precision on the UTA-RLDD dataset.(3)Embedded deployment of fatigue detection systemIn this thesis,the NVIDIA Jetson Nano embedded board was selected as the hardware platform for porting the face detection,facial landmark detection,and fatigue detection models.By taking advantage of hardware acceleration,the entire system achieved a real-time performance of 17-18 frames per second(fps),meeting the realtime operation requirement.In addition,the accuracy of fatigue detection on the embedded platform remained consistent with the results obtained on a computer platform.The accuracy on the embedded board reached 86.6%,with a precision of 86.8%.