Dissertation Feature - Based Reconstruction of 3 D Primitives from Multiple Views Joachim

The modeling i.e. the capturing of geometric information plays a key role for our fast growing urban communities. Until recently the process of capturing three-dimensional data was a labor intensive task, but with the advent of high resolution digital imaging instruments a high degree of automation is within reach. Urban architecture however, proves to be a very challenging territory for image-based modeling methods. Nevertheless modern photogrammetric and computer vision methods feature both, high robustness to cope with complex outdoor scenes and improved efficiency to allow the processing of huge amounts of data in reasonable time frames. This work presents a collection of methods for the efficient feature-based 3D modeling of urban environments. High resolution digital images are the sole data source. The term feature-based modeling in this context means that the proposed methods do not directly work on the pixel-based image information, but higher level geometric features are extracted in an initial preprocessing step. Every subsequent method then operates on those primitives. The topics presented span low-level feature extraction such as edges and ridges and corners via the robust detection of mid-level features such as ellipses and 2D line segments and new efficient methods for extracting 3D primitives from 2D features. The main contributions of this work are methods for extracting vanishing points, robust fitting of regular polygons, a method for the efficient matching of points-of-interest via a semi-global descriptor and finally a method for the efficient feature-based 3D reconstruction from multiple images. In the experimental section an in-depth analysis of the presented methods, concerning their robustness as well as their accuracy, is conducted. The experiments are conducted on synthetic as well as real data sets.