Registration of 3D models and video images in the affine space for augmented reality

Abstract

This thesis deals with the problems about the registration and the occlusion of virtual objects in real images and additionally the error minimization algorithm of lens distortion for AR system. Motivation of this thesis is that we want to view virtual models of new structures in campus by our AR system. The given problems are formulated from the motivation and the problems are grouped into three parts. Those were registration, occlusion, and design problem of the AR system. In the registration problem, the traditional methodologies can be divided into two groups. The groups are the camera calibration methods and the calibration-free methods. The calibration methods show accurate registration results by the internal parameters of the camera and the distortion coefficient of lens is applied to AR systems. On the other hand, the calibration-free methods show the high precision of the external parameters for registration. But the two images and non-coplanar 4 points are needed to construct the affine coordinates by the traditional methods. We propose a hybrid registration method in this thesis, only one image is needed to construct and the affine coordinates can be constructed from at least 4 coplanar points. And translation and rotation of the virtual objects are determined at any position in the affine coordinates. In the resolving methods for occlusion problem, the image flow methods in computer vision researches have been used to determine the depth information of the unknown real objects. Theoretically the relative image flows are determined from 6 measurement points in minimum condition. The image flows are obtained by the gradient methods or the matching methods and we use the gradient method. And the depth information is determined by the image flows of the segmentation results. As a result, the depth information is used for rendering of the virtual object that occluded with the real unknown objects and the usefulness of the registration method is verifie...