Building Extraction from Stereo Aerial Imagery Using Dynamic Programming

In this paper, a novel technique for building detection and extraction and simple building reconstruction from stereo aerial imagery is presented. This research hypothesises that geometric distortion in buildings will lead to occlusion at depth discontinuities. Depth discontinuities around buildings can be identified by determining the occlusion. Accordingly, the roof can be distinguished from the ground. Occlusion usually occurs in the direction that is perpendicular or at the edge angled to the baseline, and no occlusion occurs when the building edge is parallel to the baseline. Therefore, another stereo pair should be used to detect the depth discontinuities around the object. Dynamic programming algorithm is implemented to detect occluded pixels at the depth discontinuities. Accordingly, the occluded pixels can be identified in the form of a point cloud. The produced point cloud is scattered and cannot be used to identify or extract the building boundary. Therefore, the point cloud is converted to a raster image for detecting buildings that are shown as blobs. The algorithm is later used to extract the building shape and construction. The proposed technique is fully automated and does not require human interference. The algorithm is applied to two study areas. Two stereo pairs are used for the first study area, and only one stereo pair is available and applied for the second study area. Analyses show that the correctness and completeness of object accuracy assessment for the first study area are 100% and 97%, respectively; those for the second study area are 83% and 75%.