AN INTEGRATED SYSTEM ON LARGE SCALE BUILDING EXTRACTION FROM DSM

Large scale automatic building extraction is difficult in Japan, because of the high density of buildings and various roof types especially in the residential area. We propose an automatic and integrated system for building extraction from DSM, in order to model the outlines of the buildings in 2D plane. In our previous research, the building extraction based on foreground/background marker gives good result and most buildings can be detected. However, it is still difficult to model the building because the shape is not accurate enough. The active contour is a kind of segmentation method widely used in medical image processing. It merges edge detection into gray segmentation so that the contour converges at the right edge of the certain object. It can be employed locally on the DSM to obtain the outline of a building if the initial contour is given to close to the building. Thus we developed a scheme of accurately locating the building by marker controlled watershed and accurately segmenting the building by active contour. There are two basic processes in this scheme. First is the initial building segmentation, or the building location. Watershed of the gradient of the DSM can get the objects' contour. To avoid over segmentation, the markers of the terrain and the off-terrain must be labelled and set to local minimal. There are three types of markers to be labelled, including two of buildings and one of terrain. The methods of marker labelling for small houses and big houses are different. The markers of the small houses are decided by local maximum of DSM, while those of the big or high buildings are extracted using area morphology method. The terrain markers are detected by the watershed of the DSM not crossing the building markers. The marker controlled watershed segmentation on the gradient of the DSM gives the contours of the off-terrain objects. Then, each building object is used as a mask by enlarging its size, and active contour method is implemented to segment the masked DSM to obtain the subtle contour of the building. Level set as a typical solution of active contour is employed because of its fast iterating process. The contour of the level set can be converged exactly at the outline of the building for a local DSM including only one house. RGB image can be used instead of DSM, but the result is not as good as that of DSM. These methods and scheme construct a system of large scale and automatic building extraction. The experiment in Saitama, Japan is given as the evaluation of the proposed system.