Multiple User Real Time Eye Detection For Auto-Stereoscopic Display

In the last decade,especially from 2010,stereoscopic display has developed in the speed faster than that in any other time period,and different kinds of stereoscopic screen has been applied in almost all fields of our life.According to whether the stereoscopic display needs auxiliary devices or not,stereoscopic display can be divided into two kinds:non-autostereoscopic display which needs users to wear auxiliary devices and autostereoscopic display which does not need users to wear auxiliary devices.Because autostereoscopic display can be viewed directly by users without wearing any auxiliary device,making viewing stereoscopic display more comfortable and convenient,more and more research institutes and researchers pay attention to the research and development of autostereoscopic display.And autostereoscopic display has become a research hot spot.With the camera fixed on the top of the autostereoscopic display screen,eye positions of multiple users in the camera view field can be detected in real time.And then a pair of views of the same image can be projected into the eyes of every user,making users viewing stereoscopic scene.This is a very effective and very important kind of autostereoscopic display,which has high stereoscopic scene resolution and takes low transport width.Up to now,many well-known research institutes has joined the development of this kind of autostereoscopic display.This kind of autostereoscopic display needs an eye detection system which not only can detect eye positions of multiple users in real time,but also should be robust for illumination variation,face rotation,partial face occlusion,users wearing glasses,users' fast motion and other circumstances.Therefore,the eye detection system should have high detection accuracy rate,and should be real-time and robust for different circumstances.Given the detection platform of 2.8GHZ frequency and 1G memory single CPU personal compute,the eye detection system illustrated by this paper can obtain eye detection accuracy rate of 96.10 percent,and only takes less than 20 millisecond,satisfying real-time need.At the same time,the eye detection system illustrated by this paper is robust for illumination variation,face rotation,partial face occlusion,users wearing glasses,users,fast motion and other circumstances.To detect the eye positions of multiple users accurately,real-timely and robustly in every image frame,the eye detection system illustrated by this paper introduces near-infrared active illumination.Firstly,user faces are detected,and then user eyes are detected.And face tracking algorithm is introduced to improve the efficiency of eye detection.At the same time,online prediction,feature selection based on different face area,Kalman filtration and tracking are introduced to diminish the inference from face rotation,partial face occlusion,user wearing glasses and users' fast motion and improve the robustness of the whole detection system.The concrete methods concluded in the eye detection system introduced in this paper can demonstrated respectively as follows.(1)Firstly,unified Haar feature space is generated from discrete Haar feature space and classical Haar feature space;then with pyramid detection model,AdaBoost algorithm are utilized to detect faces.To restrain over-learning phenomenon,weight adjusting restrain strategy and normalization factor adaptive adjusting strategy are brought up;to reduce the time complexity of AdaBoost machine learning,CUDA calculation strategy which is based on parallel calculation is introduced;to make the detection system more robust for face rotation,online prediction strategy is introduced;to make the detection system more robust for partial face occlusion,feature selection based on different face area strategy is brought up.(2)AdaBoost classifiers and SVM classifier are combined in cascade structure to detect eye positions.Firstly,AdaBoost classifiers are utilized to remove fake eye candidate positions very quickly,and then SVM classifiers are utilized to detect eye positions accurately.To make the AdaBoost detection prestage more robust,correlation coefficient calculation algorithm is brought up;to restrain the effect of wearing glasses on eye detection,Kalman eye filtering algorithm based on template matching is brought up.(3)According to the motion feature of human face that face motion can not be in coincidence with any motion model strictly,extended Kalman forecasting algorithm is brought up.In this kind of Kalman forecasting algorithm,template matching of eye position is utilized to be state transfer model,and this kind of Kalman forecasting algorithm can predict approximately face positions and these positions will be detected preferentially,improving the efficiency of eye detection.A series of experiment are conducted on the eye detection system introduced in this paper.The face detection accuracy rate can achieve 98.2721 percent and the eye detection accuracy rate can achieve 96.1033 percent when the number of users are less than twenty,and the system processes twenty-five frames per second,satisfying the real-time standard.At the same time,experiment results show that the detection system is robust for face rotation,illumination variation,wearing glasses,partial face occlusion and users'fast motion.