Building detection from orthophotos using binary feature classification

Building detection in orthophotos is crucial for various applications, such as urban planning and real-estate management. In order to realize accurate and fast building detection, a non-interactive approach based on binary feature classification is brought forward in this paper. The proposed approach includes two major stages, i.e., building area detection and building contours extraction. In the first stage, a sequence of intersections is obtained by superpixel segmentation in the subsampled orthophoto, and then building area is reserved roughly according to the classification of intersections. In the second stage, the sequence of intersections is updated by superpixel segmentation in the building area from original orthophoto, and then building contours is extracted in accordance with the classification of intersections likewise. The local feature of the intersections is descripted employing our extremely compact binary descriptor, and is classified using binary bag-of-features. Experiments show that benefiting from binary description and making full use of texture details and color channels, the proposed descriptor is not only computationally frugal, but also accurate. Experiments are also conducted on orthophotos with different roof colors, textures, shapes, sizes and orientations, and demonstrate that the proposed approach are capable of achieving desirable results.