A Markerless Registration Method for Augmented Reality based on Affine Properties

Abstract

This paper presents a markerless registration approach for Augmented Reality (AR) systems based on the Kanade-Lucas-Tomasi (KLT) natural feature tracker and the affine reconstruction and reprojection techniques. First, the KLT tracker is used to track the corresponding feature points in two control images. Next, four planar points are specified in each control image to set up the world coordinate system. The affine reconstruction and reprojection techniques are used to estimate the image projections of these four specified planar points in the live video sequence, and these image projections are used to estimate the camera pose in real time. A primitive approach that illustrates the basic idea is first discussed and a robust method is given to improve the resistance to the noise. Pre-defined markers are not required in the proposed approach. The virtual models can still be registered on the proper positions using the proposed method even if the specified region is occluded during the registration process. © 2006, Australian Computer Society, Inc.