Driver Fatigue Detection Based On Face Recognition

Fatigue driving is one of the most important causes of traffic accidents.With the rapid development of our economic,we have more private cars and the number of fatigue driving has increased year by year.Therefore,it is significant to detect the fatigue state of the driver and send warning signal in abnormal state.Existing methods are mainly divided into three kinds of categories: fatigue detection based on physiological signals,fatigue detection based on the driver's operating behavior and vehicle trajectory,and fatigue detection based on computer vision technology.As the first two kinds are inconvenient or not real-time,there is no large-scale investment in practical application.In recent years,great achievements have been made in the fields of computer vision,artificial intelligence and hardware technologies such as CPU and GPU,thus driver fatigue detection based on computer vision has attracted lots of research interests.Because that the state of eye open and closure is the most distinguishing feature in fatigue detection,researchers usually combine eye closure detection methods and eye state analyzing algorithms to detect fatigue driving.The most commonly used eye closure detection methods are based on gray features of eye images,eye shape features,template matching or classical image features like LBP,Gabor,HOG,etc.However,most of the existing methods have high requirement of image quality and applying conditions,or the calculation process is complex,which all reduce the practicability of the algorithms.Based on technologies of face recognition and face alignment,we proposed an algorithm for driver fatigue detection.The main idea of the algorithm is to make use of the identity information and learn the classifying model of eye open and closed online.The algorithm is unsupervised and only need to execute the learning process when a new person is recognized.In most of time,the classifying model is being updated.In summary,the main work of this thesis includes the following three parts:1)A new fatigue detection algorithm is proposedExisting computer vision based driver fatigue detection technologies usually ignore the individual differences and only use a fixed classifying model which is trained offline,thus these algorithms can only be applied to the scene that is similar to the training conditions,and the robustness of the algorithms is not satisfying.We proposed an algorithm which combines face recognition and an online eye closure detection method.The proposed method make use of different face information and is more practical.2)Improvement of face recognition based on dictionary learningDictionary learning is an important area in face recognition,which has been widely studied,and there are many state-of-arts algorithms.Due to the limitation of computational complexity,it is necessary to reduce the dimension of the testing images before the process of dictionary learning.Existing algorithms regard the dimensionality reduction process as a step of preprocessing,but do not discuss the process of reducing dimension too much.However,different image dimension reduction methods will lead to different recognition accuracy.In this thesis,we propose a dynamic dimensionality reduction algorithm to minimize the information loss in the dimensionality reduction process by optimizing the dimensionality reduction matrix,and finally achieve higher accuracy of face recognition.3)An eye closure detection algorithm is proposedExisting eye closure detection algorithms can hardly meet the changing practical environment,and cannot learn the classifying model online to update the classifying model.We propose a new eye closure detection algorithm based on online low-rank subspace clustering.The algorithm can learn the specific classifying model online in real time,thus is much more practical.Experiments on commonly used datasets prove the effectiveness and robustness.