| Man-made environments are full of line segments, and a complex curve can be approximated with multiple straight-line segments. Therefore, line segment matching is an important computer vision problem, and it is a powerful tool for solving registration problems in 3D urban modeling .;Image-based 3D modeling is one of the core goals in computer vision, in which a great challenge is image matching. The approaches based on local features are currently the most effective matching techniques for wide-baseline images. The detection and description of existing local features are directly based on pixels. In the first part of the work, a novel local feature called Line Signature is introduced. Its construction is based on curves. Curves are extracted from images and are approximated with line segments. A line signature is a local cluster of these line segments. Extensive experiments have shown that wide-baseline image matching using line signatures has significant advantages over existing approaches in handling low-texture images, large viewpoint changes of non-planar scenes, and illumination variations.;Next, a robust approach is presented for automatic registration of aerial images with untextured aerial LiDAR data. Airborne LiDAR has become an important technology in large-scale 3D city modeling. To generate photo-realistic models, aerial images are needed for texture mapping in which a key step is to obtain accurate registration between the two data sources. Existing registration approaches based on matching corners are not robust, especially for areas with heavy vegetations. Our approach based on line segments has 98 percent success rate.;Airborne remote sensing technologies provide information of building rooftops but the detailed structures of building facades can only be captured from the ground level. Therefore, in order to create 3D models with high-resolution geometry and texture for both roofs and facades, it is necessary to integrate the two data sources. In the final part of the work, an interactive system is proposed that can rapidly create georeferenced 3D models of groups of buildings with high-resolution texture for both roofs and facades by integrating orthorectified aerial images and ground-level panoramas. To greatly reduce the user interaction, a semi-automatic approach is proposed for matching line segments detected in ground-level panoramas with those in orthorectified aerial images. |
