Recognition of 3D Objects in Aerial Images Based on Generic Models

Availability of actual three-dimensional data for geo-information systems has become of great importance for an increasing number of tasks. Since the acquisition of such data is mainly done with the help of semi-automatic tools so far, a large research program called “Semantic Modeling” was started 3 years ago with the aim of improving image interpretation by incorporating application domain knowledge represented by explicit models. In our sub-project we apply CLP for the recognition of 3D objects (i.e. buildings) in aerial images. Logic programming constitutes the platform for the representation of image and object models and the control strategy of the reasoning process.
Generic 3D models (constructive solid geometry (CSG), augmented by constraints) are applied to represent the large number of different building types on the one hand. Image segmentation results in features of different classes, giving a symbolic 2D image description on the other hand. In order to match object models to image data, a third kind of model (aspect graph) is used, bridging the gap between the 3D volumetric and 2D image data. Such aspect graphs are transformed to CLP clauses, and matching is done by solving the resp. CSP. Our current prototype is based on ECLIPSE and extends the built-in CLP(FD) solver to cope with complex objects.