Detecting buildings in aerial images using shape descriptors

Image segmentation and an interest operator are combined to identify characteristic regions that are likely to correspond to parts of buildings. A set of general purpose shape descriptors are calculated and each region is assigned a classification probability. The shape attributes of classes relating to buildings and non-buildings are represented by a set of training shapes, from which these probabilities are derived. The classification probabilities are used by our high-level vision system which fuses several sources of information to interpret the scene.