A Study On Recognizing Driver Fatigue And Distraction State By Means Of CNNs And LSTM

Owing to the rapid development of china's car industry and people's living standards,the per capita ownership of national automobile has gone through a huge grow,leading to the high frequency of traffic accidents.Of all the causes of traffic accidents,the improper behaviors of drivers rank more than 90%.Fatigue and distraction are the major factors within the improper behaviors.Thus it is of great importance to make an early warning through monitoring driver's dangerous state,which would in turn reduce traffic accidents.The paper mainly clarified the detection methods of driver fatigue and driver distraction,principle of related algorithms,and ways utilized in building,training and optimizing model.The research also designed the fatigue and distraction experiments based on the selected physiological signal data of drivers,developed recognition model of fatigue and distraction by way of deep learning technique,which was of great meaning in detecting emergency state.Firstly,a detailed summary of the researches on driver fatigue and driver distraction was elaborated,the defects of the original machine learning in detecting was pointed out,and a recognition model based on physiological signal of drivers collected by way of deep learning techniques was developed.Secondly,the structure and the error back propagation algorithm of convolutional neural networks and long short-term memory networks were detailed introduced.On the basis of the theories,the driving simulation experiments were designed by combining the current driver's fatigue and distraction experimental methods.During the experiments,the signals were collected and divided into fatigue set,distraction set and normal driving set,with each set grouped into training,validation and testing sets,respectively.The CNNs(convolutional neural networks)and LSTM(long short-term memory networks)models of driver's fatigue and distraction state were built separately and optimized by the visualization during the training process.Finally,by analyzing the accuracy among CNNs,LSTM and other traditional machine learning models based on supporting vector machines and Gaussian Maximum Likelihood Classifier,it was found that CNNs and LSTM show better performance than the traditional machine learning methods coupled with feature extraction.In addition,the sequence model of LSTM contains better accuracy which can reach up to 94.3% in the recognition of driver's dangerous state when comparing with that of CNNs.