3d Building Extraction from Stereo Pair High Resolution Satellite Imagery

In this paper, we will present a highly automated method of extracting 3D building information. Our approach set out to locate the vertices of building roof. For a start, we will limit our development to high rise buildings with approximately flat rooftop. If the main corners and intermediate vertices can be found, the 3D building model can be constructed. Our study is applying stereo pair satellite imageries; we first manually select a rectangular area of interest which contains our building of choice. This is the only manual process at the moment. We then compute the digital surface model (DSM) using normalized cross correlation template matching technique. Furthermore, we isolate the building of choice by discarding the non-building features according to the computed surface heights. If we just locate the edges of the DSM at this stage, we will end up with an unacceptable irregular geometry which is smaller than the actual roof and sometimes a building split up into a few parts. In order to find the correct building corner, we go back to the two images of the stereo pair to compute the edges from the pixel values. The rooftop found from the DSM is then expanded by a few pixels until edge pixels are detected. Clearly, there are still many false alarms such as when true edge did not show up as an edge pixel or edges picked up were not true roof edge. Using a series of rules, most of the false alarms were rejected and a series of roof edge pixels, but not sufficient to form the complete roof polygon, are retained. Subsequently, we developed an algorithm to search and extract the useful vertices of the building from the remaining edge pixels. Finally, the vertices with three-dimensional coordinates and the topological relations for each building are extracted and joined to form the roof polygon. The algorithm has been tested on stereo pairs of GeoEye-1 and WorldView-2 images.