Delineation and counting of buildings in aerial images

We propose a method to delineate buildings from aerial images and rapidly count the number of buildings damaged by a disaster. Traditional segmentation techniques may divide a single building into multiple structures when some parts of the building are in shadow and others are not. The proposed method first segments regions from aerial images by using a method robust to shadows. Then, point clouds of three-dimensional coordinate values are generated from stereo images, and ground regions are excluded by using these point clouds. Finally, the remaining regions are examined for possible merging with neighboring regions. The decision to merge is based on the sum of the rectangular and square indices under the assumption that rectangles and squares are the dominant shapes of buildings. The proposed method was applied to images of an area in Japan, and it was found that the merging of regions into one building was successful in many cases.