Driver's Face Fatigue Detection And Research Based On Convolutional Neural Network

Fatigue detection technology has been an important detection technology in the driver's step-by-step process,which can effectively detect and prompt fatigue driving.Current fatigue detection performs fatigue judgment based on the physical state of the driver's facial expression.Traditional facial fatigue judgment and recognition based on machine vision cannot express the relationship between various fatigue characteristics.From the perspective of input information processing,the deep learning feature is that neural networks can extract the required features layer by layer on the data.Deep learning can explore the non-linear relationship between driver fatigue and facial fatigue characteristics.Therefore,this paper proposes a fatigue detection method based on neural networks.First,this paper proposes a non-invasive fatigue detection method based on driver facial expression analysis.This method uses multiple facial expression parameters to determine the driver's fatigue level.The method first uses the Viola-Jones face detection method to ensure the driver's presence in the video.The lips are searched by spatial fuzzy C-means(S-FCM)clustering.The pupil state detection is performed in the upper part of the face detection window.The state of the mouth and eyes is transmitted to the fuzzy expert system(FES)to judge the fatigue state of the driver.This method has strong generalization ability for different users,and the judgment is accurate and reliable.Second,a picture of driver fatigue judged by the fuzzy expert system(FES)is used as a training sample for fatigue confidence,and a CNN neural network is trained using the fatigue confidence sample.CNN neural network can extract the driver's facial non-linear fatigue,and train the CNN neural network fatigue detection model by kernel convolution processing image hidden information and inverse tuning.Secondly,in actual driving operations,due to the influence of the driver's attitude,lighting changes,and background changes,there are great challenges for face detection and localization.We propose a deeply cascaded CNN(MTCNN)neural network framework,which is a deep recurrent network divided into three phases.In addition,during the learning process,we propose a new online hard sample mining strategy,which can automatically improves performance when it is unnecessary to manually select sample.On challenging FDDB and wider face detection benchmarks and AFLW face alignment benchmarks,our method verifies the accuracy of the method while maintaining real-time performance.Finally,the performance of the neural network is tested in the laboratory and the real driving environment.The specific realization conditions are combined with the fatigue judgment,and the future research directions are prospected.