Investigation Of Driver Fatigue Detection Method Based On Intelligent Processing Of Physiological Signals

Fatigue driving is one of the primary causes of road traffic accidents,which endangers social and public safety as well as human life and property.As a result,it is critical to investigate the detection method of fatigue condition during driving in order to reduce fatigue driving and improve the traffic environment.Physiological signal as a mapping of human physiological changes can objectively respond to human mental state,and detect both sensory fatigue and non-sensory fatigue at the same time.Among many physiological indicators,the EEG signal has been regarded as the "gold standard" of fatigue detection methods,and the EOG signal also contains various eye movement information.Therefore,this thesis uses deep learning methods to detect driver fatigue based on the complementary information of EEG and EEG signal features.The study focuses on the feature extraction,fusion method and different feature selection methods,and other concerns.The main contents include:(1)EEG signal and EOG signal feature extraction.Initially,the Short-Time Fourier transform is applied to the pre-processed EEG signals,and then the frequency bands are divided with 5 band and 2Hz band resolution.The power spectral density and differential entropy characteristics were extracted,respectively,and the features were smooth-processed to obtain the fatigue EEG feature model.Independent component analysis and minus rule were used to separate the EOG signals,after that,continuous wavelet change and peak detection were used to extract eye movement features such as blinking,saccade and fixation.(2)Fatigue driving detection model of one-dimensional convolutional encoder.Based on the stack autoencoder structure,the 1D convolutional network layers and the maximum pooling layers were used to replace the fully connected layer to construct the 1D convolutional autoencoder.The extracted EEG and EOG features were taken as the inputs of the encoder,and the deep features were classified to obtain the model performance.The root mean square errors and correlation coefficients for multiple subjects reached 0.10 and 0.88.In addition,this paper also studied and compared the effects of different feature combinations on classification results.(3)Time-Space-Frequency multidomain fatigue driving detection model based on multimodal signals.To address the problems of inadequate feature extraction and low recognition accuracy of driving fatigue detection methods,this paper combines convolutional autoencoder and recurrent neural network to perform fatigue detection from time-space-frequency multidomain features.Through the cross-training and testing of the model,the root mean square error and correlation coefficient reach 0.08 and 0.96.The proposed model has small standard deviation across multiple subjects and the best fatigue recognition performance compared with other models.This thesis also illustrates that the fatigue detection model with multimodal feature fusion outperforms the model performance using EEG or EOG alone.The algorithm model of driver fatigue detection based on multimodal physiological signals studied in this paper provides a new method for intelligent warning of fatigue driving,and the research results have significant scientific and practical implications for autonomous driving,medicine health,and brain-computer interaction.