Stable online natural feature-based registration algorithm for AR system

An efficient approach of real-time camera registration applied for augmented reality systems was presented, which purely relied on a single camera for visual tracking of the nature features extracted from the real environment. Giving a small number of calibrated key-frames and a rough 3D model of the part of the real scene, camera pose could be accurately achieved by matching inputting image and the key-frame, whose viewpoint was as close as possible to the current inputting one. Wide baseline correspondence problem was solved by rendering intermediate image. Extended Kalman filter was applied for jitter correction. Algorithm performance was tested by real image sequences. Experimental results demonstrate that the registration algorithm not only is accurate and robust, but also can handle significant aspect changes.