Building Segmentation using an Active Contour Model

This paper addresses the problem of building detection in high-resolution (1m) images and proposes an Active Contour Model as one of the solutions to ease the problem. In building segmentation, it is important to identify approximated building regions, i.e., the regions where buildings are likely to exist. Once the approximated building regions are identified, the problem of building extraction is decomposed into tracing or detecting the exact building boundary (or the boundary of a block of buildings) in each such region. If this remains unsolved, building extraction has to deal with a whole scene where there are noises and non-building features with various shapes. Often additional information, such as height information gathered from an altimeter or stereo analysis, is used to isolate approximated building regions from a scene. However, the acquisition of such information may not be available at all times. In this paper, we describe an algorithm to identify approximated building regions without additional information and hence achieve building detection in a more robust manner. For the isolation of such regions we used an Active Contour Model, which has been “actively” used in medical images to segment “smooth” boundaries. Since our problem domain involves segmentation of the features that are not smooth, we modified a Balloon model which is one of Active Contour Model to enable to “trace” recti-linear building boundaries. In our approach, building detection carries out in the following procedures. Firstly, lines are extracted from an image; secondly, using this line information the modified Active Contour Model is applied to the image to identify approximated building regions; and finally building boundaries are extracted by grouping lines in smart manner around the approximated building regions. We tested this approach with aerial photographs digitized at 1m ground sampling distance. We selected the resolution to simulate the performance of our approach with high resolution spaceborne imagery, which will be available in the near future. Test results support that an Active Contour Model can be used to identify approximated building regions and that our approach can detect buildings in recti-linear shapes only using the information that can be extracted from a single scene.