Driver Fatigue Detection Based On Deep Learning

An unsupervised feature extraction and state discrimination method was proposed to detect driver's fatigue and solve the contradiction between the exponential growth of data and the expensive cost for preparing corresponding labels for machine learning based on the image data set.The fatigue detection system was divided into two functional modules,facial feature extraction and fatigue state detection.The first part uses an improved convolutional autoencoder in unsupervised learning which is of efficient connection,and the space features for images will not be damaged.The idea of ??greedy layer-by-layer training is applied to the convolutional autoencoder to form a stacked convolutional autoencoder,as a result the model training speed was improved and the premature network degradation was avoided,which helps to lead to find a better numerical solution of parameters.The visualization of the extracted facial features and the classify results verified the effectiveness of this feature extractor.Next,the k-means clustering with improved initializing mean vector method,which was determined by predicted category probabilities of Softmax classifier,was used to classify the category.Using the contour coefficient to evaluate the clustering effect,the initialization method proposed in this paper was improved by 0.35 compared to random initialization,and the states recognition results of drivers' images also proved the effectiveness of the classify method.Finally,the experiment of end-to-end fatigue state recognition results on the testing set was performed on the entire system.The prediction results were evaluated by the confusion matrix,and the detection accuracy rate of the fatigue state reached 87.5%.Besides,the gradient descent method was used to update the parameters in the feature extractor and K-means clustering algorithm consistently,causing our system can be implemented in the same iterative optimization framework and the program smoother.