to
Or
nt
Gruen, Armin
SEMI-AUTOMATED APPROACHES TO SITE RECORDING AND MODELING
A. Grün
Federal Institute of Technology (ETH) Zürich
Institute for Geodesy and Photogrammetry
CH-8093 Zürich
agruen(a)geod.baug.ethz.ch
Working Group V/3
KEY WORDS: Object recognition, Modeling, Reconstruction, Semi-automation, Topology
ABSTRACT
Automated object extraction is a key issue in photogrammetry and computer vision. For the delivery of precise, reliable
and complete results human interference is indispensable. Therefore the semi-automated extraction techniques have
gained much interest recently. This paper gives an overview about the status of this technology, with special emphasis
on the work and results of our own group. We will limit our discussion to the reconstruction of outdoor scenes, and here
in particular to the generation of buildings and complete 3-D city models. Both aerial and terrestrial applications will be
covered.
“...the idea that extracting edges and lines from “ It must not be overlooked that segmentation is
images might be at all difficult simply did not occur actually one of the central and most difficult
to those who had not tried to do it. It turned out to be practical problems of machine vision. "
an elusive problem.”
Marr, 1979, p.16 Davies, 1990, p. 77
1 INTRODUCTION
Site recording and modeling has been an important topic in photogrammetry from its very beginning in the middle of
the 19" century on. Since then technologies have changed several times fundamentally. Today the issue of full
automation of all processes involved has led to widespread research activities in both the photogrammetry and the
computer vision communities. However, progress is slow and the pressing need to produce precise, reliable and
complete datasets within reasonable time has had scientists and developers turn towards semi-automated approaches
(compare Figure 1). While the tasks may differ in terms of required resolution (level of detail) of models, type of
product (vector model, hybrid model, including mapped texture, attributed model with integrated thematic information),
size of dataset, sensor platform (satellite, aerial, terrestrial), sensor and data type (images in various forms, laserscans,
scanned maps, etc.), one common problem remains in all cases and that is the automated extraction of objects from
images. A typical example is the task of automated building detection and reconstruction, which is difficult for many
reasons.
The most common source of data are 2-D images which lack direct 3-D information. Aerial images may differ from
each other with respect to scale, spectral range of recording, sensor geometry, image quality, imaging conditions
(weather, lighting), etc. Objects, like buildings, can be rather complex structures with many architectural details. They
may be surrounded by other disturbing man-made and natural objects. Occlusion of parts is common and the
geometrical resolution may be limited. Therefore the corresponding images are of very complex content and highly
unstructured. Solving the problem of building detection and reconstruction under these conditions not only is of great
practical importance but also provides an excellent testbed for developing image analysis and image understanding
techniques. The basic problem in object extraction stems from the fact that automated image understanding is still
operating at a very rudimentary level. This applies both to close-range and aerial/space applications. However, there is a
remarkable relation between image scale and successrate in extraction. At smaller image scales the level of geometric
modeling becomes lower and the image context is easier to grasp since the relationships between objects are less
distorted by artifacts. Thus the extraction of DTMs and objects like buildings, roads, rivers, landuse elements, etc.
becomes less complicated.
Since users tend to request more and more highresolution results in 3-D modeling, the pressure is on to pay more
attention to the semi-automated techniques.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 309