Research On Driver Fatigue Recognition Based On Facial Features

The high incidence of road traffic accidents in China has caused huge economic losses.Research shows that fatigue driving is an important cause of traffic accidents.This article focuses on the real-time fatigue driving recognition on the embedded platform,and selects fatigue recognition method based on facial features which is non-contact and easy to realize.This method is composed of face detection,facial landmark detection and fatigue recognition,which is a progressive task method.The main research contents of this article are as follows:1.Research on Driver Face Detection Algorithm.In order to realize real-time face detection on embedded platform,a number of improvements are proposed based on MTCNN algorithm: using Deep Separable Convolution replace CNN,using Full Convolution Network generate candidate box,using Global Average Pooling replace Full Connection,using NonMaximum Suppression algorithm based on Intersection over Union,exploring the optimal values of two key parameters of Image Pyramid,etc.Through these improvements,a lightweight face detection network is designed.The detection accuracy of this network reaches the level of mainstream algorithms,and the detection speed is greatly improved.2.Research on Driver Facial Landmark Detection algorithm.In order to realize facial landmark detection and head pose estimation simultaneously,a heterogeneous facial landmark detection network based on multi-tasking learning is designed.The network avoids the problems of training difficulty and low accuracy of traditional multi-tasking network,and uses Bottleneck simplify the network structure.A Cost Sensitive Loss function is designed to solve data imbalance.The network has high accuracy,high speed and excellent comprehensive performance.3.Research on Fatigue Recognition Algorithm based on Facial Features.Eyes closing,yawning,and frequent nodding can be described as facial fatigue features according to the three indicators of PERCLOS,mouth closing degree and Pitch Angle.A facial fatigue feature vector is constructed for each frame,and then the vector is stitched into a spatio-temporal fatigue feature sequence for a video.A fatigue recognition network based on LSTM was designed to recognition the spatio-temporal feature sequence,and the network was optimized by frame skipping detection and multi-scale feature fusion.The algorithm has good accuracy,and the structure is simplified which can realize real-time fatigue identification on embedded platform.4.Design of Fatigue Recognition Software.In order to improve the usability and enrich the function of the algorithm,a driver fatigue recognition software is developed based on NVIDIA Jetson TX2 embedded platform.The software has beautiful interface,simple operation and complete functions.In addition,a management system is developed,which realizes the function of fatigue driving data transmission and terminal management through network.