Research On Vehicle Driver Fatigue Detection Method Based On Multi-feature Fusion

Fatigue driving is one of the important causes of traffic accidents.Therefore,the development of a driver fatigue detection system that can detect the driver fatigue state in real time is of great significance to the improvement of automobile safety.In order to solve the problems of the existing driver fatigue detection methods,such as low detection accuracy,poor robustness,and insufficient consideration of the change laws in time series,a method based on machine vision and multi-feature fusion is used to study the driver fatigue detection method in the actual driving environment.The main research contents of this article are as follows:(1)Driver effective facial area extractionConsidering the actual driving environment,a cascade of regressors based on ensemble of regression trees is established to detect key points of the driver's face,and the driver facial feature extraction methods are proposed based on the key points of the driver's face.First,a weak cascade regressor is built based on tandem regression trees,and several weak cascade regressors are used to form a strong cascade regressor for detecting key points of the driver's face.At the same time,the error function of node division in the regression tree is expanded to solve the problem of missing labels in dataset.Then,based on the positional relationship between the key points of the driver's face,the overall features of the driver's face,eye images,and mouth images are extracted.(2)Driver facial behavior perceptionBased on deep learning,a model for detecting the features of the driver's face is established.The model uses a deep neural network to roughly process the features of the driver's face,and uses the results as reference information for driver fatigue detection.A deep compressed convolutional neural network model to analyze the driver's eye/mouth state is established.Based on the multi-channel convolutional neural network,the model adds residual connections to improve the performance of the deep model.The deep model compression strategies greatly reduces the amount of parameters and calculations without significantly affecting the detection accuracy,witch improves the detection efficiency of the model.A model for driver gaze zone estimation was established.The model uses the distance and angle relationship between key points of the driver's face as the driver's head posture estimation features,and calculates the grayscale of the driver's eye slice after binarization to obtain the driver's pupil position.The model uses the distance and angle between the key points of the driver's face as the features of the driver's head posture estimation.The calculation of statistical grayscale is performed on the driver's eye slices after binarization to obtain the position of the driver's pupil.Finally,the driver's head posture features and pupil position are used as feature inputs,which are fed into random forest to estimate the driver's gaze zone.(3)Establishment and optimization of multi-feature fusion modelA multi-feature fusion model based on time series is established.Based on long short-term memory and bidirectional recurrent neural network,a bidirectional long short-term memory that can integrate driver's multi-feature is established to detect driver fatigue.The model takes the detection results of the driver's facial features,eye state,mouth state and gaze zone as input,and the output is a two-class classification of fatigue state,which fully considers the change law of driver fatigue state in time series.Based on the proposed multi-feature fusion model,several deep model optimization strategies are analyzed and selected,such as parameter initialization,adaptive learning rate,batch standardization,hyper-parameters selection and representation learning.(4)Building driver fatigue state detection modelBased on the detection results of the driver's eye state,mouth state and gaze area,fatigue characteristics such as PERCLOS,the maximum duration of eye closure,the duration of the mouth state,and the duration of the driver's sight away from the main area were calculated.Based on the fatigue characteristics,the performance of the driver's fatigue detection model based on single feature is analyzed,and compared with the driver's fatigue detection model based on multi-feature fusion.