Urban Object Detection Using a Fusion Approach of Dense Urban Digital Surface Models and Vhr Optical Satellite Stereo Data

In this paper we describe a new approach for the extraction of urban objects from very high resolution (VHR) optical stereo satellite imagery. Such data is delivered from sensors like Ikonos, QuickBird, GeoEye or WorldView-II. These sensors provide ground sampling distances (GSD) of 0.5 to 1 m for the pan chromatic channel and 2 to 4 m for the multispectral channels. Normally good digital surface models (DSM) can only be expected at 1/3 to 1/5 of the original GSD. But we present a new approach which uses the generation of dense disparity maps based on computer vision approches and fuse these disparity maps with additional information gained from the original imagery to allow the extraction and afterwards modeling of urban objects. This can be achieved due to the fact that the generated disparity maps are constructed on one of the original images. So a direct pixel to pixel correlation of the height (represented by the disparity) and the spectral information (represented by the pan sharpened original image) can be done. Applying methods for the generation of a digital terrain model (DTM) which represents the ground without elevated objects and spectral classification allows the separation of typical urban classes like buildings, trees, roads, low vegetation, water and so on. These classes will be treated individually in the modeling step to generate a simplified 3D model of the observed urban area. The results are presented, compared to the original imagery and discussed.