Acquiring structure from motion requires a solution to multiple
problems in the field of computer vision. We present a general pipeline for the
reconstruction of 3D models from multiple calibrated images and the
implementation of it's stages. The main stages are feature extraction, feature
matching, camera pose estimation and finally scene reconstruction. The
selection of homogenous regions as features is based on the assumption that
they mostly represent planar surfaces of the imaged scene, which allows us to
represent mappings of such regions between images with homographies. In the
search for feature matches is computationaly more efficient to transform the
regions into a normalized form and represent it with a vector descriptor. The
subsequent region normalization technique improves the accuracy in the stages
of feature matching and camera pose estimation. A novel method of multiple
homography decomposition is used to obtain the camera poses and scene structure
allowing us to create a reconstruction of the imaged scene. The presented
methods are analyzed using synthetic data.
|
