Real-time Eye Positioning Algorithm For Stereoscopic Display

The positioning and tracking of human eyes is the key module of auto-stereoscopic display,and the performance of the module has a direct impact on the user's viewing experience.By obtaining the locations of user's eyes quickly and accurately and using advanced optical devices,our system makes it possible for users to view stereoscopic content without any auxiliary equipment.At the same time,eye positioning technology is an important part of intelligent technology which has many applied directions and a bright future.The hierarchical processing method is applied in this paper.Firstly an improved elliptical model is applied to detect skin color in YCbCr space,and the candidate areas are selected through integral image and the coverage of skin color.Then the area of eyes are located by the two-layer(face-eye)classifier,of which the face classifier is composed of 11 strong AdaBoost classifiers and the eye classifier is composed of 3 strong AdaBoost classifiers and one SVM classifier.Precise pupil positioning is carried out when the location information of eyes is obtained.Several pupil positioning methods are tested,and the improved hybrid projection algorithm is raised up after the comparison in the aspects of effect and efficiency.A motion-related filtering algorithm is presented to overcome the irregular jitter of the positioning result.The filter can determine if the current positioning result is valid in real time according to the correlation between positioning results and tracking results.Meanwhile,in the front stage of face detection,Kalman prediction and the dynamic switching between detection frames/tracking frames are applied to accelerate the processing and improve the overall efficiency.In the training part of AdaBoost classifiers,training samples with reasonable distribution are used,and the weights of weak classifiers are adjusted according to the property of human face's grayscale distribution.The original training program based on CPU platform is replanted to the GPU platform,and the training speed is significantly promoted by introducing the latest GPGPU technology.The classifiers have good robustness,especially for the large area occlusion and various expressions.Of course,there are some aspects which should be improved in our system.Our future work includes:the enhancement of positioning accuracy and success rate,improving the adaptability of various illumination and improving the training algorithm and process.Moreover,at present,the system can only handle frontal and slightly tilted faces,so how to handle faces with more postures in real time is the next problem to be solved.