Fatigue Recognition Based On Face Videos

Driver fatigue is a major factor in the causes of automobile accidents.In recent years.Traffic transport project of our country is developing quickly, but the malign traffic transport is also ascending. 78.5 percent of deaths by all kinds of accidents is caused by traffic accident during 2002. The traffic accidents caused by driver fatigue took part about 20 percent, which exceeded 40 percent of the most serious traffic accidents. It is reported that traffic accidents caused by driver fatigue take large percent in America, Japanese and England and so on. It can be seen that driver fatigue has become the main hidden trouble of traffic accidents. Therefore, detecting driver fatigue is extremely important to improving transportation safety.Driver fatigue monitoring technologies can be divided into subjective methods and objective methods. Subjective methods get driver fatigue levels from subjective questionaries, note tables made by drivers, questionaries about the habit of sleep, and Stanford measurement tables of sleep. Objective methods detect driver fatigue from the parameters about drivers and cars. Research on fatigue monitoring can make the measures for the prevention of dangers caused by driver fatigue more convenient and direct. In the past decade, more and more attention has been paid to the technology research and equipment development for fatigue monitoring. We can divide the methods for driver fatigue monitoring into three categories: physiological parameters, individual characters of drivers and performance of cars under driver control based on the information source. When people are in fatigue state, their circadian characters vary much. Circadian signals have some metabolic rules. Investigators get people's circadian signals such as electroencephalogram, cardiogram, electro-oculogram, electromyogram and so on through med equipment to detect drivers'fatigue in real time. Detecting fatigue based on drivers'characters is to get drivers'videos through cameras, analyze the videos using computer optic method and judge whether drivers are in fatigue state or not according the results of the characters information. When drivers are in fatigue state, the veracity, agility and smartness of their actions all will be influenced, so it is feasible to detect drivers'fatigue by how the cars are controlled by drivers.When a driver fatigues, he will take many special visual cues on his face. In this paper we present a novel framework based on computer vision for driver fatigue detection that combines various visual cues typically related to fatigue. We combine visual cues from mouths and eyes systematically which characterize eye closed and yawning to infer driver fatigue.Eyes'changes are the most important character to reflect peoples'fatigue state, when man is tired, his eyelids stir frequently and eyes close for longer time. Judge eyes'state through detecting eyes'character seen as a charater to detect fatigue. AdaBoost is used to extract the most discriminative features from the LBP features of eye areas and constructs a highly accurate classifier to get the eye visual cue. LBP arithmetic operator is not parameterization which describes local dimensional structure. LBP is a strong tool to describe texture. Because of strong texture-distinguish ability and smaller compute cost, LBP has been used in pattern recognition widely. Basic LBP arithmetic operator converts the 3*3 pixels around every pixel of the image into binary numeric according to the grey level of the centre and takes the result as a binary digit. LBP has 256 possible values, so LBP codes can denote 256 LBP modes, and every value denotes a LBP local mode. In this article, we use LBP8,2 arithmetic operator to distill character of every pixesl and compute the LBP value of every pixel, the original eye image and LBP image then come into being.The article uses Adaboost arithmetic to choose the LBP character of eyes and train strong categorizer. Adaboost arithmetic is the most popular Boosting arithmetic, and the AdaBoosting arithmetic advanced by Freund and Schapire gives a very simple and effective study method to choose character and not-linearly classfy.AdaBoost has achieved much success in detecting face and other application areas. AdaBoost is an ensemble study algorithm. It creates a series of weak categorizers. AdaBoost conclude advanced arithmetic and basic arithmetic, basic or weak arithmetic is actually to choose suitable discriminated function, advanced arithmetic is a iterative process, it chooses the weak categorizer according to their capabilities in the training gather, then a strong categorizer forms by the linear combination of weak categorizers. Yawning is also an important cue of driver fatigue. Fusion the cues of eye closed and mouth yawning can get a more accurate and robust fatigue detection. In the paper, to get the mouth cue, we propose to use gray projection to locate the left and right mouth corners, extract the texture features by Gabor wavelets, and apply LDA to classify the mouth features to detect yawning. When drivers yawn, the texture of the mouth has great change while the geometric feature changes. The texture of the mouth includes more information. Gabor wavelets are widely used in computer vision, texture analysis and object recognition. The two dimensional Gabor wavelets can get the best resolution in spatial field, temporal field and direction. In the experiments, we use a discrete form of the Gabor wavelets which include 5 frequencies and 8 orientations. The image convolves with the real and imaginary parts of the Gabor wavelets at the left and right moth corners. Then, at each moth corner, we get 40 complex Gabor wavelet coefficients. Through the Gabor transformation, we can represent the texture of the left and right mouth corners with a vector of 80 elements. The purpose of the Fisher linear discriminant is to find a feature projection space which minimizes the ratio of the within class scatter of the projected samples to the between class scatter. That is to say the samples of different classes are separated optimally i.e. the more the scatter between classes is, the better it is while the samples from the same class are congregated i.e. the less the scatter within the classed is, the better it is. So, the samples can be separated correctly in the projection space.The analysis and understanding of fatigue based on single character have some localization, and the dubious factors may influence veracity and dependability of some fatigue characters. Therefor, many researchers fuse many characters information to analyze fatigue state. Analyse of fatigue state through fusing many characters information is to distinguish whether fatigue brings or not through two or more of blink, staring direction, gesture of head, ship of mouth, and expression of face and so on. In the article, information amalgamation occurs in the decision-making level. The result through matching of eyes'clue and mouth'clue composites a character vector which is input into LDA, and the last decision-making comes from the matching of the LDA's output and the stencil.The method advanced by the article is tested by ten people in real life, which contain changes of gestures and illumination. The result indicates that visual fatigue clues of the article have stronger dependability and veracity.