Automatically Constructing the Third Dimension of IMGeo

Abstract

The past year a new Dutch standard for the registration of large scale topography was released, called IMGeo 2.0. This standard includes descriptions and attributes for the majority of objects seen in public spaces. A possible extension to IMGeo 2.0 allows all objects to be modelled in 3D, based on object descriptions and techniques from the CityGML standard. This thesis explores the feasibility of automatically constructing such 3D models by combining 2D geometries and 3D height information. Numerous methods for this type of construction already exist, differing in objects reconstructed, amount of detail, restrictions on construction approach, assumptions on input data and requirements on output data. The approaches created in this thesis aim to construct 3D models in a robust way, accepting sub-optimal results as long as they are a decent and valid representation of reality. Four different construction methods were designed. The first builds large 2.5D terrain models based on point triangulation. These 2.5D surface models form the basis of the actual 3D models. The second and third method construct horizontal roof 3D block models of buildings, with one approach dividing the roof based on point clusters and the other using repeated square subdivisions. The last and most complex method tries to construct 3D building models with full roof shapes using plane and line detection techniques. Degrees of success vary across these methods, with lower success rates the more complex the models become. The constructed terrain model has a number of improvements over a standard triangulation, but terrain generalization is really a field of its own and a lot more improvement can be accomplished. The methods for block models proved to be robust and performed quite well, but can be improved with respect to roof border accuracy. The method that constructs full roof shapes is not robust enough yet, but shows promising possibilities. Overall some creative ways to automatically construct 3D models have been designed, either entirely from scratch or as a modification of existing work, combining techniques from amongst others the fields of geometric algorithms and image processing.