Secondary abstract: |
A wide range of applications using 3D building models exists; such as computer games, city marketing, disaster management, tourist information systems, simulations of noise propagation and surveillance of sustainable construction. Complete acquisition of large urban scenes has become feasible using multi-aspect-oblique-view ALS; however, automated generation of detailed 3D models, the main focus of this thesis, still poses a significant challenge.
To enable enrichment of a 3D building model with windows, the 3D wire-frame building model and the ALS point cloud are first automatically co-registered. The novel approach to window extraction presented in this thesis exploits evidences about window positions in processed oblique-view ALS point cloud and façade image textures. Laser beam penetrates glassy window areas, and thus points found behind a segmented façade plane, projected onto the façade plane, give reliable evidence about intrusion positions. On the other hand, high values of gradient on a texture are usually due to window frames. These two facts are exploited when extracting initial window patches. Additionally, binary masks, obtained by region growing of homogeneous parts of façade textures, are used to eliminate certain façade artefacts and to improve shape of window patches. The assumption, that many windows of the same kind are on the same floor, is used for the refinement procedure. First, façade textures are divided into horizontal blocks, representing floors. Second, a search for non-similar window patch templates within each block is performed. Third, to obtain additional window patch positions, the chosen templates are cross-correlated along the respective block.
Eleven façade planes of an existing 3D wire-frame building model are textured with extracted patches, representing windows and other intrusions. Despite different arrangements of windows, varying window sizes, and relatively strict evaluation method, the method results in 63% detection rate. What is more, the method is mostly data-driven and the detection rate outperforms the method using only oblique-view ALS (Tuttas & Stilla, 2013). The windows are well defined, since the basis for most of window patches are connected components of edges belonging to window frames. |