Co-registering and Normalizing Stereo-based Elevation Data to Support Building Detection in Vhr Images

Building detection from very high resolution (VHR) remote sensing images has been an active research area for more than two decades. This is because building information is crucial for analysing urban environments, and also because VHR images are the ideal geo-spatial data source for extracting and mapping such information. Since optical images are a 2D projection of the real world surface, elevation data provide an additional key component for detecting and delineating buildings. Therefore, optical images and elevation datasets must be accurately co-registered in order to detect and map buildings. This is a challenging task by itself. Additionally, the co-registered elevations need to be normalized (i.e. removing the terrain elevation) so that elevation values only represent the heights of aboveground objects. The normalization process often introduces elevation errors and reduces the building detection quality. To remedy this problem, this paper introduces an integrated method for co-registration and terrain elevation filtration to directly generate a set of 3D points that facilitate detecting elevated buildings in a 2D image. An image-space co-registration of a dense set of image matching points is proposed. These points are identified and matched in a pair of stereo VHR images and then assigned their corresponding elevations generated by photogrammetric/aerial triangulation. Once an accurate image-space registration of the triangulated elevations is achieved, an image classification technique is executed to identify the terrainlevel natural and artificial objects (e.g., grass and roads). The associated terrain-level elevations are then extracted and utilized in the normalization process that allows a direct elevation-based building detection. The proposed method was executed and validated. Improved building detection results over traditional methods were achieved.