Model and context-driven building extraction in dense urban aerial images

This paper presents a method to extract buildings in monocular urban aerial images without prior knowledge of illumination. With a building concept model to interpret buildings at different levels and scales, building extraction is carried out in two stages: sunshine parts extraction and self-shadow parts extraction. Based on region-oriented radiometric features, images are first simplified and segmented into three parts: sunshine parts of high objects, shadow regions and sunshine ground. To verify initial segmentations, a method of estimating the direction of shadow cast was proposed by a shadow context model. Building extractions were then refined by context and a modified partial snake model with the aid of the shadow cast direction, thus sharply reducing the iteration complexity and the influence of illumination. An innovative method for extracting self-shadows on gable roofs is also discussed.