Driver Drowsiness Detection Method Based On Facial Expression Features

Driver drowsiness is one of the major causes of fatal traffic accidents. Byprocessing and analyzing the driver’s facial images based on machine vision which is anon-contact, reliable and real-time method, drowsiness can be effectively detected. Theproblems of the existing studies are: indicators, mainly statistics indicators, are notsufficient enough to describe the expression information; indicators are limited to facelocal characteristics, such as the eye movement characteristics, lack of overall facialexpression feature. The detection accuracy will improve by enriching the facialexpression features with futher mined facial expression information and simulating thehuman intelligent understanding process. This thesis proposed a new way of descriptionand simulation of drowsy expression, outspreaded the study in extraction of drowsyexpression features and development of driver drowsiness detection algorithm. Thestudy includes:Firstly, facial drowsy expression characteristics and the expressivness of staticfacial images were analyzed. The mechanism and elements of fatigue expression, aswell as the characteristics of state variation process were qualitatively analyzed. Byexploring the difference between the human ability on fatigue states distinguish viadynamic videos and images, feasibility of fatigue detection using static facial expressionwere verified.Secondly, a new drowsy expression describing method based on geometricconstruction was proposed, and the feature validity was vertified. A series ofdrowsiness-related facial feature points were defined, and a feature quantization methodbased on triangular mesh structure was proposed; distribution characters of indicators indifferent drowsiness levels as well as the corresponding relation with intuitive facialexpression were explored, and their significance of difference was tested.Finally, drowsiness discriminant algorithm based on expression features wasestablished. A multi-layer tree classifier was designed with indicators filtration andlinear discriminant function in each layer, to discriminate three types of fatigue ofdrivers.Experimental validation result shows that, with the facial morphological features asindicators, the detection accuracy based on static facial images is78.5%. Discrimination index fully embodies the facial fatigue expression features. The algorithm accuracy ishigher than the average of human fatigue discriminant, reached almost expert’s fatiguestate judgment level.