EXTRACTION OF RECTANGULAR BUILDINGS IN AERIAL IMAGES Samuel Vinson , Laurent D. Cohen , and Fr´ ed´ eric Perlant EADS Matra Syst` emes et Information, 6, rue Dewoitine, 78142 V´ elizy-Villacoublay, France svinson, fperlant @matra-ms2i.fr CEREMADE Universit´ e Paris IX Dauphine Place du Mar´ echal de Lattre de Tassigny, 75775 Paris cedex 16, France cohen@ceremade.dauphine.fr ABSTRACT This paper introduces an exhaustive process for assisting buildings extraction out of the Digital Elevation Model (DEM) and the orthoimage. We focus on rectangular buildings, which are the most common constructions. The method is a region-based approach for extracting above-ground structures and for estimating them with rect- angles. We establish a relation between the eigenvalues of the covariance matrix and the rectangle dimensions. The Hausdorff measure is used to validate this estimation. The estimated rectangle shape may be neither well lo- calized nor well sized. We use a parametric model for im- proving the estimation. The final rectangle estimation is used to make a precise rendering of the 3D reconstruction of the scene. Keywords: Building extraction, Rectangular estimation, Deformable Templates, Aerial image 1. INTRODUCTION In the civil and military fields (infrastructure of telephony, impact studies...), it is increasingly necessary to use realistic models. For examples, authors of [5, 2, 7] have presented automatic methods for terrain modeling (ground, vegeta- tion, and buildings...) based on aerial or satellite photographs. From this modeling and aerial images, the orthoimage, i.e, the vertical view of the scene, is computed (Fig. 1a). Several works intend to improve building rendering in the Digital Elevation Models (DEM). Brunn et al. [3] min- imize a criterion based upon rules of collinearity and right angle. Lin et al. [11] use perceptual grouping to aggregate building edges. Vestri’s last work [14] improves accuracy DEM by modifying the method of generation by correlation especially on the building frontages. Lee et al. [10] present a semi-automatic system to gen- erate 3D models with rectilinear hypotheses. This system attempts to minimize the time and the number of user in- teractions by defining rules to substract or add rectangles to models. In our approach, we want to minimize the operator work- load. Above-ground structure extraction (vegetation and constructions) (Fig. 1b) is carried out on the DEM by the algorithms presented in [4, 1]. Section 2 deals with estimating the rectangle parame- ters over any given blob. In section 3, we define a criterion for checking the estimation efficiency. A method which im- proves the rectangle estimation is introduced in section 4. In the following section, we present an algorithm for splitting complex buildings in several rectangles. At last, we give an overview of this whole process through examples. 2. ESTIMATION OF RECTANGULAR PARAMETERS We intend to model each above-ground element by a rect- angle. The rectangle center of mass and its principal axes are those computed over the element itself. We make an es- timate of the two remaining rectangle parameters by means of the element set of points. The center of mass is given by the first order moment (Eq. 1). The principal axis is computed through the estima- tion of the second order moments (Eq. 2, Eq. 3). and (1) (2)