Research On Fatigue State Detection Method Based On Facial Key Feature Fusio

As the problem of road traffic accidents caused by driver fatigue driving is increasing day by day,it is particularly important to develop a fatigue detection system with high accuracy and generalization to reduce the incidence of traffic accidents.Aiming at the problems of poor accuracy,poor robustness and intrusiveness of wearing special signal acquisition equipment,which cause great interference to the driver’s normal driving,a non-contact fatigue detection method based on key feature fusion is proposed in this paper.The specific work is as follows:First,add a illumination compensation module.Aiming at the problem that the traditional fatigue detection method does not consider the adverse effect of light changes on subsequent detection during actual driving.The histogram equalization is selected as the illumination compensation algorithm in this paper,and it is found that the image quality is greatly improved after illumination compensation.Secondly,an improved MTCNN(multi-task convolutional neural network)face detection network is proposed.Due to the problems of large model space,complex calculation and low detection accuracy in traditional face detection algorithms,this paper reduces the complexity of the model by improving the original network structure.Aiming at the problem that the original network generates too many image pyramid layers and the detection time is too long,two improvement measures are proposed:optimizing the minimum face size and compressing the number of image pyramid layers.The results show that the improved network not only effectively reduces the input data of the model,which can greatly reduce the computational complexity,but also greatly improves the detection efficiency under the premise of ensuring the detection accuracy.Then,multi-feature extraction.Aiming at the situation that a single fatigue feature is prone to low accuracy or failure of the algorithm due to the driver wearing glasses and sunglasses,this paper proposes a multi-fatigue feature fusion algorithm for eyes,mouth and head.The mouth and eyes feature extraction is realized by two algorithms.The first one uses the self-built data set to train the self-built network EMSD-NET(eyes and mouth state detection net)to realize the mouth and eye opening and closing state recognition.The accuracy is 99% and 98% respectively through the verification of the test set;the second based on the transfer learning strategy uses Res Net-18 as the backbone network to retrain 68 facial key points in the YTF dataset,and calculate the eye aspect ratio(eye aspect ratio,EAR)and the mouth aspect ratio(mouth aspect ratio,MAR)indirectly realize the recognition of the state of the mouth and eyes.The N-point perspective algorithm is used to realize the head pose estimation.Finally,multi fatigue parameter fusion based on rough set theory.The extracted key features are further refined into six fatigue parameters: blink frequency,percentage of eyelid closure(PERCLOS),eye closing time,yawning times,nodding frequency and nodding time.The current driver’s fatigue state is comprehensively judged by using rough set theory.Through experiments,it is found that the algorithm is non-invasive,will not interfere with the driver’s driving,and the cost is low;The detection accuracy is higher than the traditional detection algorithm;It can effectively overcome the influence of complex environment such as light,face occlusion and wearing glasses.When a certain feature fails,it can still work normally.Compared with the single feature detection algorithm,it has strong robustness.