基于改进的深度卷积神经网络的人脸疲劳检测

Face Fatigue Detection Based on Improved Deep Convolutional Neural Network

针对疲劳驾驶检测问题,提出一种以softmax损失与中心损失相结合的深度卷积神经网络算法。首先,利用含有方向的梯度直方图(histogram of oriented gridients,HOG)和级联分类器(support vector machine,SVM)算法的Dlib库中预训练的人脸检测器,来检测驾驶员的脸部区域。其次,使用级联回归(ensemble of regression trees,ERT)算法实现脸部68个关键点标定及眼睛和嘴巴的定位。最后,为了优化softmax损失在深度卷积网络分类中出现的类内间距大的问题,加入中心损失函数,提高类间差异性、类内紧密性以及驾驶员脸部疲劳状态识别准确率。在自建测试集和YawDD哈欠数据集中的实验结果显示,该方法能够准确地识别检测驾驶员疲劳表情,平均识别准确率达到98.81%。与传统的疲劳驾驶检测识别方法相比,该方法可以自动进行疲劳特征提取,并且训练准确率、检测识别率及鲁棒性得到提高;与未改进的深度卷积网络相比,检测识别的概率平均提高了约5.09%。

Aiming at the problems of fatigue driving detection,a deep convolutional neural network algorithm was proposed.The algorithm wascombining softmax loss and center loss to detect the facial fatigue state of the driver.Firstly,a pre-trained face detector in Dlib library containing histogram of oriented gradient(HOG)and support vector machine(SVM)algorithm was used to detect the presence of driver’s faces.Then,68 key points in driver’s face plus eyes and mouth were located through the ERT(ensemble of regression trees)algorithm.Finally,in order to optimize the problem that the sortmax loss had large intra-class spacing in deep convolutional network classification,the center loss function was introduced to optimize and improve the difference of inter-class,the compactness of intra-class and the recognition accuracy of the driver’s facial fatigue state.The experimental results on the self-built test set and the YawDD yawn data set demonstrate that the method can accurately identify the driver’s fatigue state,and the identification average accuracy rate of our algorithm is about 98.81%.The method can automatically extract fatigue features and improve training accuracy and test recognition rate compared with the traditional fatigue driving detection algorithm.Robustness is significantly improved.The recognition accuracy increases by approximately 5.09%compared with the unimproved deep convolution neural network.