Resumen de Algorithms for the reconstruction, analysis, repairing and enhancement of 3d urban models from multiple data sources
Over the last few years, there has been a notorious growth in the field of digitization of 3D buildings and urban environments. The substantial improvement of both scanning hardware and reconstruction algorithms has led to the development of representations of buildings and cities that can be remotely transmitted and inspected in real-time. Among the applications that implement these technologies are several GPS navigators and virtual globes such as Google Earth or the tools provided by the Institut Cartogràfic i Geològic de Catalunya.
In particular, in this thesis, we conceptualize cities as a collection of individual buildings. Hence, we focus on the individual processing of one structure at a time, rather than on the larger-scale processing of urban environments.
Nowadays, there is a wide diversity of digitization technologies, and the choice of the appropriate one is key for each particular application. Roughly, these techniques can be grouped around three main families: - Time-of-flight (terrestrial and aerial LiDAR).
- Photogrammetry (street-level, satellite, and aerial imagery).
- Human-edited vector data (cadastre and other map sources).
Each of these has its advantages in terms of covered area, data quality, economic cost, and processing effort.
Plane and car-mounted LiDAR devices are optimal for sweeping huge areas, but acquiring and calibrating such devices is not a trivial task. Moreover, the capturing process is done by scan lines, which need to be registered using GPS and inertial data. As an alternative, terrestrial LiDAR devices are more accessible but cover smaller areas, and their sampling strategy usually produces massive point clouds with over-represented plain regions. A more inexpensive option is street-level imagery. A dense set of images captured with a commodity camera can be fed to state-of-the-art multi-view stereo algorithms to produce realistic-enough reconstructions. One other advantage of this approach is capturing high-quality color data, whereas the geometric information is usually lacking.
In this thesis, we analyze in-depth some of the shortcomings of these data-acquisition methods and propose new ways to overcome them. Mainly, we focus on the technologies that allow high-quality digitization of individual buildings. These are terrestrial LiDAR for geometric information and street-level imagery for color information.
Our main goal is the processing and completion of detailed 3D urban representations. For this, we will work with multiple data sources and combine them when possible to produce models that can be inspected in real-time. Our research has focused on the following contributions: - Effective and feature-preserving simplification of massive point clouds.
- Developing normal estimation algorithms explicitly designed for LiDAR data.
- Low-stretch panoramic representation for point clouds.
- Semantic analysis of street-level imagery for improved multi-view stereo reconstruction.
- Color improvement through heuristic techniques and the registration of LiDAR and imagery data.
- Efficient and faithful visualization of massive point clouds using image-based techniques.
