| The National Center for Statistics and Data Analysis reports that fatigue is one of the leading factors in car accidents,and that if drivers could react 0.5 seconds earlier before a crash occurs,it would reduce crashes by 60 percent.Therefore,the development of driver fatigue detection technology is crucial,which can effectively test whether the driver has fatigue driving and reduce the incidence of car accidents.In order to solve the shortcomings of current fatigue detection methods with low accuracy or poor real-time performance,this paper integrates HOG face detection algorithm and KCF target tracking algorithm and introduces deformable convolutional neural network for state recognition of extracted eyes and mouths,and the main work of the study is as follows.(1)The acquired images are pre-processed for the negative effects of light and noise on fatigue detection.In this paper,we use histogram equalization processing to enhance images collected in natural environments with complex lighting.Noise removal is achieved using median filtering in nonlinear filtering to reduce the negative impact of noise on fatigue detection.Image pre-processing improves the identifiability of the inspection equipment imaging and the accuracy of the inspection system.(2)The kernel correlation filter tracking algorithm is introduced to improve the traditional face algorithm for the problem of low accuracy.The HOG face detection algorithm and KCF target tracking algorithm are fused to perform fast tracking of the detected faces and extract continuous and stable target faces to achieve more efficient extraction.After extracting the driver’s face image,this paper calls the Dlib library for face key point extraction,which increases the accuracy of the algorithm.(3)Deformable convolutional neural networks were introduced for state recognition of the extracted eyes and mouths.The proposed deformable convolution operation is an extension of the standard convolution operation that uses an additional standard convolution operation to introduce a spatial offset of the values on the grid points compared to the standard convolution for changing the location of uniform grid samples,allowing the convolution operation to learn free deformation information and improving the accuracy of state recognition.A head pose algorithm is introduced for real-time detection of the driver’s head,and this obtains information about the driver’s head posture.(4)A multi-feature fusion fatigue detection method is proposed based on the states of the eye,mouth and head.Compared with current fatigue detection algorithms,the proposed new method can detect driver fatigue status in real time with high accuracy.In this paper,there are two ways to verify whether the driver is driving fatigued,through the self-contained camera for real-time detection of the driver’s status,and also through the pictures and videos captured by the camera to be imported into the fatigue detection system designed in this paper.The fatigue detection method based on multi-feature fusion designed in this paper has good robustness as verified by the collected data sets and simulated driver states. |
