Research On Multi-Feature Fatigue Driving Detection Method For Image Based On Support Vector Machine

Fatigue driving detection can be used to prevent the occurrence of traffic accidents, which ensures the security of drivers' life and assets. The traditional fatigue driving detection methods is on the base of the driver's eye states, which is a single feature for fatigue evaluation. The proposed method is based on the state of the eyes with the fusion of head pose, and the improved support vector machine classification is used. Compared with the traditional method of fatigue driving detection, the method proposed in this thesis has many characteristics and higher credibility, the main work is as follows:(1) Three commonly used color spaces are studied and YCbCr color space is used to detect the skin color for its better color matchings AdaBoost is used for the face detection in the region where the skin color was detected. Harris corner detection algorithm is used for eye positioning and state detection in the region of detected faces. Compared with the traditional eye state detection algorithms, high efficiency, detection speed are obtained, and the influence of eyelashes for the closed state of eyes is avoided,(2) The traditional support vector machine (SVM) is improved by using K-type kernel function and logistic kernel function to substitute the traditional kernel function, which makes the learning and generalization capability improved and better classification performance obtained.(3) The improved support vector machine is applied to fatigue judgment. Compared with the traditional eye-based fatigue judgment methods, the method in this thesis combines the head attitude, which makes reliability and accuracy rate higher.The effectiveness of fatigue driving detection with 3 different types of kernel functions in SVM are compared. Experiment results indicates that the mixed kernel function that combined K-type kernel function with logistic kernel function used in SVM can get better results compared with the traditional SVM methods. At the end, some shortcomings of this method are summerized and some advice on how to improve it is given.