Driver Fatigue Detection Based On Facial And Human Behavior Characteristics

The number of road accidents caused by driver fatigue has been increasing year on year.Accurately assessing the fatigue status of drivers and providing reminders can to some extent reduce traffic accidents caused by driver fatigue and reduce the incidence of accidents.Due to their single feature information,the predictive ability of existing fatigue detection methods in practical usage scenarios is easily affected by the driver’s driving actions.In order to meet the robustness requirements of fatigue detection,this article establishes multiple feature representations related to fatigue based on the facial information,head posture,and human posture of the driver while driving.Multiple types of fatigue features are used as inputs,and the fatigue features are fused and classified through multiple model fusion methods to achieve the detection of the driver’s fatigue state.The main work of the study is as follows:(1)A driver fatigue detection method based on face inverted pendulum model and information entropy is proposed.Firstly,the Practical Facial Landmark Detector(PFLD)is used to extract the coordinates of 108 key points on the driver’s face and calculate the pitch angle,yaw angle,and roll angle that represent the head posture.Then,the key point coordinates are used as input to establish the face inverted pendulum model and calculate the kinetic energy and potential energy of its link system.The kinetic energy,potential energy and head posture data of the inverted pendulum model are taken as fatigue characteristics,and the information entropy value of fatigue characteristics in a period of time is calculated using the sliding window.Finally,the information entropy is classified based on the Long-short Term Memory network model to achieve the classification and prediction of driver fatigue.(2)A fatigue detection method based on human posture model and degree of dispersion has been proposed.Replace the backbone residual network(Res Net)in the simple baselines of human posture with the mobile network(Mobile Net-V1),and then perform deconvolution processing.Extract the driver’s human posture information from the video sequence,establish and calculate multiple fatigue related features,use sliding windows to calculate the degree of dispersion based on human posture fatigue features,introduce temporal factors into the fatigue prediction process,and finally use the degree of dispersion as the input of the classifier to identify the driver’s fatigue state classification and prediction.(3)Based on a modified Stacking multi-model fusion classification method,the fatigue features of the face,head and human posture are used as inputs for the final driver fatigue detection.Considering the differences in data analysis and training principles among different classifiers,multiple machine learning algorithms are used as base classifiers in Stacking multimodel fusion to leverage the advantages of each model.In the process of fusion classification,the dataset is weighted and fused.Compared with individual feature fatigue detection algorithms,multi-model fusion can more effectively extract driver fatigue status information from different types of fatigue features,and accurately detect driver fatigue status in usage scenarios.The result of fatigue feature classification based on facial and head posture information reached 94.17% through validation of the dataset collected through simulated driving.The accuracy of fatigue feature classification based on human posture is 95.21%.Using all fatigue features as inputs for Stacking multiple model fusion,the classification result was ultimately improved to 97.26% after combining multiple feature information.The improvement of classification results further demonstrates that the fatigue detection method based on multi feature fusion designed in this paper can effectively fuse multiple fatigue feature information of drivers.The proposed method can effectively improve the detection level of driver fatigue status,and has important theoretical significance and engineering application value.