Feature-Based Reconstruction of 3D Primitives From Multiple Views

The work presents a system for the efficient modeling of buildings from images and 3D laser scanner data. Aiming for a preferably simple polygonal representation the focus lies on the use of prominent image features typically ocurring in images of man-made structures. For the modeling process a set of facade images acquired from ground level and a dense 3D point cloud from a 3D laser scanner serve as source data. Extracted image features such as edges and line segments are used to compute characteristic building attributes like vanishing points and points-of-interest that provide useful information for an automatic image orientation process. Vanishing points also provide cues about the principal directions of the depicted facades. The orientation of the images is based on the assumption of the presence of prominent facade planes visible in the images. From an oriented image sequence 3D line data are extracted by an automatic line matching algorithm. The extracted 3D lines provide an accurate description of discontinuities and borders occuring in the facade planes. A sweep based procedure detects half-planes attached to the generated 3D lines. The boundaries of the half-planes are found by mutual intersection of different half-planes and incorporating dense 3D point clouds coming from laser scanners or an image matching process. By extracting those 3D points that are lying close to the plane a rough outline for the plane is detected and refined by a sweeping process. not assigned GG: Internal Project of the Institute