This habilitation presents a collection of papers dealing with the automatic extraction of buildingsfrom Airborne Laserscanner (ALS) data, supported by aerial imagery. Building extraction consistsof two stages: the detection of buildings, essentially a classification task, and the geometrical recon-struction of buildings in previously detected regions of interest. Both stages are dealt with in thiswork. First, a rule-based method for building detection is presented. This method can use both ALSdata and multi-spectral information in the form of a normalised difference vegetation index (NDVI).This method can be applied in a hierarchical framework of coarse generation of a digital terrainmodel by morphological filtering. The second method for building detection presented in this workis based on the Dempster-Shafer theory for data fusion. It uses a heuristic model for the distributionof evidence to the classes of the classification process. A thorough evaluation of that method hasshown that this model is appropriate and that most of its parameters can be determined relativelyeasily from "meaningful" entities such as a minimal building height or the approximate percentageof trees in a scene. It was shown that buildings larger than about 120 m2 can be reliably detectedusing ALS data of a resolution of 1 m and an NDVI image. Buildings between 50 m~2and 120 m2canstill largely be detected. The major influence of the NDVI was a reduction of false positive detectionsof buildings smaller than 100 m~2 by up to 15%. Building reconstruction as presented in this work starts with the extraction of roof planes fromthe ALS data. After that, a classification of the mutual geometrical relations between neighbouringroof planes is carried out, with the aim of determining the boundary polygons of these roof planes.This includes a method for the precise location of step edges in ALS data. In this process, deci-sions are based on statistical tests rather than on simple thresholding operations, thus increasing therobustness of the approach. These tests require rigorous modelling of the stochastic properties ofthe geometric entities involved. The roof boundary polygons can be grouped to form polyhedralbuilding models. Finally, the parameters of these polyhedral models are estimated in a consistentparameter estimation process that considers geometrical regularities. In this way, building modelswith a planimetric accuracy in the range of the original point spacing and with a height accuracy inthe range of a few centimetres can be generated. However, the quality of the results is limited by thesensor resolution, since the planar segmentation requires a certain minimum number of ALS pointson each plane of the roof.
展开▼