atics,
Stephan Heuel
TOPOLOGICAL AND GEOMETRICAL REASONING IN 3D GROUPING FOR
RECONSTRUCTING POLYHEDRAL SURFACES
Stephan Heuel, Wolfgang Förstner* and Felicitas Lang*
* Institut für Photogrammetrie, Universität Bonn
: Nussallee 15, D-53115 Bonn,
e-mail: Stephan.Heuel|Wolfgang.Foerstner@ipb.uni-bonn.de
* Institut für Photogrammetrie und Ingenieurvermessungen,
Universität Hannover
Nienburger Strasse 1, D - 30167 Hannover
email: felicitas.lang 9 ipi.uni-hannover.de
KEY WORDS: Building Extraction, 3D Grouping, Reasoning, Topology, Uncertain Geometry
ABSTRACT
We are developing a system for reconstructing polyhedral surfaces from multiple images. This process can take advantage
of the topological relations of the observed image features triggering and therefore speeding up the grouping of features
to polyhedral surfaces. Exploiting the statistical properties of features when grouping them leads to consistent decisions
being invariant to numbering and choice of coordinate system and allows statistical testing. This simplifies the choice of
thresholds to the definition of a scene independent significance level. We decribe the topological and statistical models
used within our system. Experiments with synthetic and real data prove the feasibility of the approach.!
1 INTRODUCTION
Reconstructing polyhedral surfaces from multiple images is a classical task in Computer Vision. In case of controlled
environment solutions are quite far advanced. However, in outdoor environments, e. g. when tracking vehicles or when
reconstructing buildings, systems have to cope with quite a number of difficulties, such as a non-optimal feature extraction
producing cluttered image descriptions with qualitative and quantitative errors. The technique of grouping significant
features to high-level structures is often used to overcome some of these problems. The rules for grouping features are
either very general and do not apply to real imagery or are rather specific and depend on the specific task. Most of the
work was done in grouping two dimensional features but there also exists some work in 3D, for an overview see (Sarkar
and Boyer, 1993).
Apart from early work (cf. (Clowes, 1971), (Brooks, 1987), (Herman and Kanade, 1986)), grouping straight lines or
planes in 3D mostly appears in the context of building extraction from aerial images (cf. (Roux and McKeown, 1994),
(Henricsson, 1996), (Frere et al., 1997) (Baillard et al., 1999)). These approaches, though embedded in a specific ap-
plication appear to be the most general ones by restricting to polyhedral surfaces. The work of (Roux and McKeown,
1994) contains the most explicit use of 2D and 3D connectivity between features. (Baillard et al., 1999) generate plane
hypothesis by half-plane detection and group 3D lines within these half-planes using collinearity and coplanarity criteria.
Furtheron they close the half-planes by plane intersections.
All these systems primarily aim at finding at least one path from the basic features to object descriptions of generic nature
and in this respect give valuable rules for grouping. However, due to the complexity of the objects and the need to use
domain specific knowledge, no general grouping rules have been established.
We are developing a system for reconstructing polyhedral surfaces in outdoor environments aiming at exploiting as much
generic knowledge as available from the structure of polyhedra, the imaging and the feature extraction process. There are
two types of knowledge: a) neighborhood relations between the atomic features, points, edges and faces, resulting from
topology, and b) crisp form relations at the object, especially planarity, and during the imaging process, resulting from
geometry. On one hand, our feature extraction procedure (Forstner, 1994) was motivated by the need to exploit all features
and their mutual neighborhood relations. In (Rothwell et al., 1996) topology as basic information in object representation
! This work has been founded by the European Community as part of the ESPRIT project “Image Processing for Automatic Cartographic Tools
(IMPACT)", No. 20.243. ue
*This work originates from the time the author belonged to the Institute for Photogrammetry at the University of Bonn
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 397