Research On Key Technology Of Vision-based Augmented Reality Registration

Augmented reality (AR) utilizes visualization techniques to superimpose virtual information on the living real world,in order to improve people perception of the external world and has. been known as one of the world’s top ten technologies used to change the world in the next20years. The key to a successful AR system is accurate, fast and robust registration,which means the perfect fusion of virtual objects with real scene for the purpose of ensuring the user experience.This paper mainly focuses on the registration issues of AR system based on computer vision and image processing,and proposes several registration methods and new ideas under different application conditions. The main tasks are as follows:Firstly,the principle of vision-based AR registration is studied systematically. A complete registration parameter derivation and parameter optimization method are proposed based on stepwise analysis of the camera imaging model and distortion model. Then the paper deeply summarizes the characteristics and development direction of the current realization of two AR modes(marker recognition and natural feature tracking registration).Secondly, an improved sub-pixel corner locating algorithm with higher corner detection accuracy and better distortion robustness is proposed for AR marker-based registration. A marker system based on Hamming coding is also designed and successfully used in AR registration. The results show that the method can effectively meet the desired speed and stability requirements of the AR system.Finally, the marker-less registration technology based on efficient tracking is a research emphasis. An improved tracking framework based on Bayesian classifier is proposed with compressed sensing theory. A random measurement matrix followed with a Gaussian distribution is bulit for sparse representation of the target features, to achieve tracking by learning and updating of positive and negative samples. The improved tracking method is applied to a kind of web-based AR registration framework with natural feature tracking. The successful marker-less positioning and rendering experiments show that this idea is feasible, which in a sense provides a theoretical basis and technical reserves for the development and popularization of the future AR applications.