Monocular Right-Angle Building Hypothesis Generation in Regularized Urban Areas by Pose Clustering

Abstract

By decomposing the building extraction problem into three sequential steps (selection, indexing and correspondence), the complexity of monocular extraction from aerial imagery can be reduced. This paper focuses on the first step, selection, during which an image subset likely to contain a single building is extracted. Pose clustering can be one way to achieve this. Based on vote accumulation, pose clustering offers the advantages of reduced complexity and a false alarm rate prediction capability. This paper describes a voting scheme for right-angle flat-roof buildings, from which image space building location hypotheses may be efficiently generated for regularized urban areas. The proposed scheme incorporates weights, constraints and uncertainties that should be implemented due to the nature of aerial imagery. Additionally, based on an occupancy model, a random vote accumulation and threshold analysis of the proposed voting scheme is presented. The main limitation of the proposed scheme is the need for a scaled building model prior to the hypothesis generation phase. Results from simulated and real imagery show that the building detection percentages as well as the quality percentage are well within range, even without a verification process.