Research On Driver Fatigue State Based On Deep Learning And Information Fusion

With the rapid and stable development of social economy in China,the number of cars is increasing day by day,followed by the soaring road traffic safety accidents,which seriously endangers the safety of people's lives and property.Fatigue driving,as one of the main inducing factors of traffic accidents,has become a hot research topic at home and abroadIt can reduce the risk of traffic safety to a great extent to detect the driver's fatigue state in time and to effectively remind or intervene when he is tired.At present,the detection of driving fatigue by machine vision is the most promising detection method in the field.However,due to the complex driving environment and individual differences of drivers,there are still some shortcomings in this method.Aiming at the above problems,based on the existing research,combined with image recognition and physiological signal processing,this paper proposes an information fusion driving fatigue state recognition method based on deep learning.The specific research contents are as follows:(1)In order to get the experimental data close to the real driving,the experimental data acquisition system is built,including the simulation driving module and the data acquisition module,respectively collecting the video data and pulse data of the simulation driver in the fatigue state and the normal state.In order to make the experimental data more reliable,the fatigue state excitation link is added to simulate the fatigue state of the driver.In the experiment,the postgraduate students were selected as the data collection objects.They were in good physical condition and had no major medical history.Finally,322 groups were collected in fatigue state and 403 groups in normal state.(2)For a single video signal source,the fatigue state of the collected data is identified based on CNN.The frame rate of the collected video data is read.The extracted image information is recognized based on OpenCV and the human eye part of the image is extracted,normalized and the image format is modified according to the network standard input.Then,the new network model is trained based on the introduction-v3 network model,and the fatigue state of each frame is identified.Finally,different experiments are conducted to compare the difference.The effective parameters are obtained from the recognition rate under the threshold value,and the fatigue judgment is made for every 10 seconds video sequence,and finally the single signal source driving fatigue recognition rate is obtained.(3)BP neural network is used to identify the fatigue state of a single pulse signal source.Firstly,the signal is preprocessed,including frequency reduction and resampling of the pulse signal segment,then the noise and baseline drift of the pulse signal are removed based on the db5 wavelet filter,22 features in the time domain and frequency domain of the signal segment are extracted,and the feature selection is based on C-SVM-RFE,and finally the importance order of the features is obtained.Finally,the fatigue state recognition model is built based on BP neural network,and the change of recognition rate under different dimension feature input after feature screening is analyzed.(4)Based on the weighted voting method,integrated learning and fusion of two signal sources are used to identify driving fatigue state.In order to further improve the recognition accuracy,the sample data is balanced based on smote algorithm to ensure that there is no deviation between positive and negative samples.The recognition rate of image signal source is mapped,and the average recognition rate is used as the weight of two signal sources for weighted voting fusion.Compared with the recognition rate of single signal source and information fusion based on relative majority voting algorithm,the algorithm in this paper has higher robustness and recognition rate advantage.