Acquisition of 3D models from a set of 2D images

The acquisition of accurate 3D models from a set of images is an important and difficult problem in computer vision. The general problems considered in this thesis are how to compute the camera parameters and build 3D models given a set of 2D images. The first set of algorithms presented in this thesis deal with the problem of camera calibration in which some or all of the camera parameters must be determined. A new analytical technique is derived to find relative camera poses for three images, given only calibrated 2D image line correspondences across three images. Then, a general non-linear algorithm is developed to estimate relative camera poses over a set of images. Finally, the presented algorithms are extended to simultaneously compute the intrinsic camera parameters and relative camera poses from 2D image line correspondences over multiple uncalibrated images. To reconstruct and refine 3D lines of the models, a multi-image and multi-line triangulation method using known correspondences is presented. A novel non-iterative line reconstruction algorithm is proposed. Then, a robust algorithm is presented to simultaneously estimate a model consisting of a set of 3D lines while satisfying object-level constraints such as angular, coplanar, and other geometric 3D constraints. Finally, to make the proposed approach widely applicable, an integrated approach to matching and triangulation from noisy 2D image points across two images is first presented by introducing an affinity measure between image point features, based on their distance from a hypothetical projected 3D pseudo-intersection point. A similar approach to matching and triangulation from noisy 2D image line segments across three images is proposed by introducing an affinity measure among 2D image line segments via a 3D pseudo-intersection line