Detection Of Driver Drowsiness Based On Convolutional Neural Network

With the increasing standards of citizen living and mertial condition,it is more comman to see sedans in our daily life.In addition,the number of commercial vehicle is also rising rapidly because of the development of transportation industry.Vehicles convenient people's daily life,but the enormous loss of life and property caused by traffic accidents are inestimable.It has been proved that the accidents caused by driver fatigue take up a large proportion in traffic accidents accroding the statistic data.Therefore,researches on driver drowsiness have been taking seriously by reserchers and government.Fatigue state determined by detecting facial feature such as eyes closure though camera is the principal method of study.However,two problems,affection of light and trade-off of accuracy and real time,are needed to deal with.A real-time and accurate algorithm which has a good adaptability to affection of light is proposed here after some works based on the two problems has been done.The method of fatigue detecting combined with multiple algorithms including digital image dealing,computer vision and pattern recognition which is utilized for extracting ocular characteristic,convolutional neural network used for recognizing the state of eye accurately and fatigue determinatoin with the combination of PERCLOS and blink frequency,follows the principle of compressing detective area gradually.The main contents and efforts of this paper can be listed as follows:1.Methods of reference white and normal histogram are chosen to compensate illumination for digital images;median filter algorithm is used for image denoising.In addition,some other methods such as image graying are utilized for image preprocessing.Image preprocessing improves the identifiability of the equipment imaging,facilitates the operation of the subsequent steps,and enhances the adaptability of the illumination influence and improves the accuracy of the detection system2.The Adaboost algorithm is used to detect face area instantaneously,as a result,the initial location would cut in half down on detection range of key points.Face classifier based on Adaboost could be trained by Haar-like feature.3.The method of ensemble regression is applied to train the facial key point detector.We could detect the facial 68 key points in area which has been detected by a face classifier mentioned in the previous step.And then,driver's eye area could be calculated by eyes location information that has been extracted from the acquired key points information.Next,we use eye aspect ratio(EAR)to determine whether the eye is open or not and capture automatically the images of eye area with state of unfolding or folding from the video stream into the corresponding classification data sets.4.Traning the recognition CNN model of eye condition with the datasets obtained by previous step and detecting driver drowsiness in real time with the determination based on the parameters of PERCLOS and blink frequency.Simulation of driver drowsiness detective algorithm here is operated in Pycharm develop environment with the programming language of Python and computer vision library of OpenCV.