A Markov-Random-Field Approach for Extracting Straight-Line Segments of Roofs From High-Resolution Aerial Images
MetadataShow full item record
This paper proposes a method for extracting groups of straight lines that represent roof boundary sides and roof ridgelines from high-resolution aerial images using corresponding airborne laser scanner (ALS) roof polyhedrons as initial approximations. Our motivation for this research is the possibility of future use of resulting image-space straight lines in several applications. For example, straight lines that represent roof boundary sides and precisely extracted from a high-resolution image can be back-projected onto the ALS-derived building polyhedron for refining the accuracy of its boundary. The proposed method is based on two main steps. First, straight lines that are candidates to represent roof ridgelines and roof boundary sides of a building are extracted from the aerial image. The ALS-derived roof boundary sides and roof ridgelines are projected onto the image space, and bolding boxes are constructed around the projected straight lines while considering the projection errors. This allows the extraction of straight lines within the bounding boxes. Second, a group of straight lines that represent roof boundary sides and roof ridgelines of a selected building is obtained through the optimization of a Markov random field-based energy function using the genetic algorithm optimization method. The formulation of this energy function considers several attributes, such as the proximity of the extracted straight lines to the corresponding projected ALS-derived roof polyhedron and the rectangularity (extracted straight lines that intersect at nearly 90 degrees). In order to validate the proposed method, four experiments were accomplished using high-resolution aerial images, along with interior and exterior orientation parameters, and available ALS-derived building roof polyhedrons. The obtained results have shown that the method works properly and this will be qualitatively and quantitatively demonstrated in this research.