Research On Detection Of Fatigue Driving Based On Deep Learning

Driver fatigue driving will affect the normal driving of the vehicle,and in seri-ous cases will threaten the life safety of the driver and passengers.Therefore,fatigue driving detection and forecasting of the driver can effectively guarantee people’s safety.researchers mainly proposed three different solutions for the problem of fatigue detec-tion:based on driver physiological feature detection,based on vehicle motion feature detection and based on driver facial feature detection.Among the three solutions,al-though the detection method based on the driver’s facial features has the advantages of non-invasiveness,high accuracy and easy operation,the blink detection method for the driver can most directly reflect the fatigue state in the detection of facial features.Therefore,the method of using blink detection to determine fatigue state is widely used in scientific research.This thesis has conducted a lot of research and in-depth analysis of various do-mestic and foreign fatigue detection algorithms.Based on previous work,mainly two related studies have been carried out:Firstly,for the problems of poor robustness and low accuracy of traditional methods,this thesis proposes the blink detection method based on deep learning has greatly improved the robustness and accuracy.Secondly,for the problem of low image exposure in low light environments,this thesis proposes a night blink detection algorithm based on low light enhancement Low-light enhance-ment has been introduced into the field of fatigue driving detection and has achieved very good results.Finally,in order to verify the effectiveness of the model,this thesis builds an experimental platform of fatigue driving detection system to facilitate the re-lated experimental verification.The main innovations and work contents of this thesis are as follows:1.In order to solve the shortcomings of traditional methods,such as low anti-interference and low accuracy,this thesis proposes a blink detection method based on deep learning.The traditional method for the extraction of eye parts is generally to per-form face detection and then locate the eyes.This method is not only time-consuming and complicated.However,this thesis directly uses the face keypoint detection network to process the image,which can perform the two tasks of face detection and locate eyes at the same time,which can meet the needs of real-time.Secondly,based on the com-mon convolutional neural network,this thesis combines the two strategies of residual learning and jump connection to improve the detail expression ability of the model,which can classify the state of the human eyes.2.In real life,when the light intensity is weak at night,the driver has more fatigue driving.The above model and the existing related detection algorithm cannot deal with the lighting problem,resulting in a low accuracy during nighttime detection.Therefore,this thesis proposes a night blink detection algorithm based on low light enhancement.In the algorithm designed in this thesis,the original image is no longer directly input into the face keypoint detection network for detection,but the original low-exposure image is enhanced before the detection.Thereby,the accuracy of the positioning of the subsequent key point detection algorithm and the classification accuracy of the open-eye classification network are improved.This ultimately improves the performance of the entire detector in a low-light environment.3.In order to test the above algorithm and test the robustness and accuracy of the algorithm,this thesis builds an experimental platform of fatigue driving detection system.The whole system is divided into two parts:fatigue driving detection terminal and background management system,which can effectively verify the correctness of the algorithm.