3-D MODELLING OF BUILDINGS FROM DIGITAL AERIAL IMAGERY 
Jussi Lammi 
ESPA Systems Ltd. 
Tekniikantie 12, FIN-02150 Espoo 
Finland 
ISPRS Intercommission Working Group II/III 
KEY WORDS: Digital Aerial Imagery, Building Extraction, Solid Modelling, 3-D Modelling, Euler-Operators, Semi- 
Automatic Measuring, Image Matching, Line Matching. 
ABSTRACT 
Three-dimensional modelling of buildings from large-scale digital aerial images is studied. The main focus is on the 
integration of photogrammetric feature extraction and solid modelling. This combination is considered necessary to 
make the data collection for realistic three-dimensional models more efficient. The paper concentrates on solid 
modelling by boundary models. The principles of boundary models are described, and their applicability to building 
extraction is discussed. The extraction of primitives for the boundary models is done in a semi-automatic way: the user 
points to the feature of interest from one image using a mouse, and the matching procedure then finds the 
corresponding feature from other images. This paper presents a search-based method for least-squares matching of a 
line segment. Here the matching itself is done in object space. The procedure can make direct use of multiple images in 
matching. The use of search-based methods in the matching of planar faces is discussed. 
1. INTRODUCTION 2. SOLID MODELLING USING 
BOUNDARY MODELS 
Automatic extraction of buildings from digital imagery has 
been the subject of many studies in photogrammetry (see Geometric modelling techniques can be divided into three 
e.g. Gülch, 1992; McKeown and McGlone, 1993; Braun, categories: graphical models, surface models and solid 
1994). Methods have been developed to solve low-, mid- models. The main categories are further divided, e.g. 
and high-level vision processes in building extraction. according to how the mathematical modelling space is 
Initially, the research was focused on low-level defined, what the representation space is and what kind 
processes like point and edge detection. This was of representation scheme exists between the modelling 
followed by the use of high-level processes to identify space and the representation space. Generally, we 
and label the objects found with low-level operators. cannot use only one modelling technique. In this case, 
Currently scientists are looking increasingly at mid-level the geometric model of the object is called a hybrid model, 
processes like grouping and spatial reasoning (see e.g. or a multiple representation of the object. 
Fórstner, 1995). It is obvious that a functional system 
combines processes from all levels. Solid modelling techniques emphasize the general 
applicability of geometric models. Mántylà (1988) defines 
Direct comparison of the different methods for building the goal of these systems to be the capability to answer 
extraction is seldom very meaningful, because the arbitrary geometric questions automatically, i.e. without 
approach chosen is strongly application dependent. In human intervention. Solid models are typically divided 
some applications the detection of buildings is of primary into three classes: decomposition models, constructive 
interest, while in others the reconstruction of buildings is models and boundary models. In decomposition models, 
also required. For many purposes it is sufficient to fit a the solid is described in terms of volume elements. The 
building model determined by few parameters into a set of most common decomposition models are spatial 
measurements (Haala and Hahn, 1995; Weidner and enumeration and octrees. Primitive instancing, which is a 
Fórstner, 1995). Some visualizations made using simple parametric model, is also a kind of decomposition 
automatically measured models of buildings are already model of an object. In constructive models, the solid is 
quite highly evolved (Polis et al., 1995). However, the thought of as a bounded set of points in three- 
construction of detailed visualizations (Meister and Dan, dimensional space. The most important group of 
1994; Gruber et al., 1995) requires that at least some part constructive models are the constructive solid geometry 
of the extraction work is done manually, or semi- (CSG) models. Here the solid model is built from a 
automatically under human guidance (see also Streilein, composed set of parametrized primitives. Boundary 
1994). models are a generalization of graphical models into solid 
models. They are constructed from vertices, edges and 
This paper examines building extraction from large-scale faces that imitate the common way of thinking of solid 
aerial images. The objective of the study is to find objects. 
methods to integrate the reconstruction and the solid 
modelling of a building into a single process. The main This chapter gives an overview of solid modelling using 
focus is on the construction of detailed solid models of boundary models. The boundary model is chosen 
buildings. The principles of solid modelling by boundary because its basic primitives are the same as the features 
models are presented and the use of semi-automatic of building extraction. It appears that powerful tools can 
tools for image matching is demonstrated. Finally, the ^ be built to update the complex data structure behind the 
advantages and disadvantages of the approach are boundary model. A more profound presentation of solid 
discussed. modelling techniques and especially boundary models 
213 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996