e aim of
) the real
for each
| at ETH
age, able
>. We do
ms, input
) to now,
lings and
dels that
dditional
f the 3"!
'ations is
s have to
and with
cost and
source of
pects like
osaicking
buildings
al. (1995,
‚CVIU -
are very
Ich et al.,
e of them
teraction,
areas and
Hug and
cz, 1999;
erson and
Christoph Kaeser
Xia, 1997; Mercer and Gill, 1998; Aero-Sensing Radarsysteme, 2000) or hybrid sensors (Gamba and Houshmand,
1999) have been used for generation of city models and building reconstruction, sometimes with a coarse building
modelling. However, these sensors are not widespread, costly, the processing of their data is complicated and not fully
automated, are in a relatively early stage of development, plus they usually do not provide the quality and high
resolution of aerial imagery, which is necessary for checking and completing the data and other purposes. Digital
photogrammetric cameras (e.g. HRSC-A) have also been used for the generation of surface models in cities (Renouard
and Lehmann, 1999). However, the problems faced there in automated object reconstruction are similar to those with
scanned aerial film imagery.
To ease automation of aerial image analysis a priori information can be used. This includes usually data from maps, GIS
and other geodatabases. In addition to existing data, a priori information can include models, constraints about the
objects and the scene, rules etc. Thus, the terms knowledge-based, model-based, rule-based etc. have been coined. A
particular difficulty in exploiting existing geospatial data for image analysis is that their accuracy and completeness are
often not sufficient, or even unknown and almost always inhomogeneous. Thus, mechanisms of estimating and
propagating the uncertainty of the input data have to be developed. Examples of approaches that incorporate a priori
knowledge for object extraction are given: (a) for buildings in Baillard et al. (1999), Haala and Brenner (1999), Stilla
and Jurkiewicz (1999), (b) roads in van Cleynenbreugel et al. (1990), Plietker (1994), de Gunst (1996), Bordes et al.
(1997), Vosselman and de Gunst (1997), Prechtel and Bringman (1998), Tónjes and Growe (1998) and (c) other more
general objects like landcover classes, urban scenes and sites in Matsuyama and Hwang (1990), Janssen et al. (1990),
Solberg et al. (1993), Chellappa et al. (1994), Maître et al. (1995), Stilla (1995), Quint and Sties (1995), Huang and
Jensen (1997), Koch et al. (1997), Roux and Maître (1997), Liedtke et al. (1997), Plietker (1997), Quint (1997a, 1997b),
Schilling and Vógtle (1997), Tónjes (1997), Zhang (1998), Walter and Fritsch (1998), Walter (1998, 1999), Growe
(1999), Kunz (1999), Tônjes et al. (1999), Pakzad et al. (1999). A second clear tendency that aims at easing automation
and improving the results is the combination of different input data that provide complimentary, but also redundant,
information and cues about the existence, shape, size etc. of an object. Terms like fusion and integration are in vogue,
although the theoretical foundations of such procedures are still not well founded, and there are no commonly accepted
methods of how they should be performed.
2 AIMS OF ATOMI
The project ATOMI is a co-operation between the Federal Office of Topography (L+T) and ETH Zurich. L+T is
responsible for map production in scales smaller than 1:25,000, generation of a nation-wide DTM, generation of digital
colour orthoimages, production of digital maps in raster and vector form etc. One of its aims is to produce a digital GIS
infrastructure (a Topographic Information System) representing a real landscape model and not a generalised
cartographic one. Important objects in this infrastructure are the transportation network and buildings. This data is often
required in different applications, especially GIS-based ones but also for visualisation, by public and private
organisations. They currently exist or are being generated in vector form by digitisation of the 1:25,000 topographic
maps by semi-automatic procedures (called VECTOR?2S5; see more details below). The aim of ATOMI is to update this
dataset fitting it to the real landscape, improve the planimetric accuracy to 1m and derive the height of the road
centerlines and one representative height for each building. For the buildings roof outline details < 2 m should be
ignored. The representative height for each building should be a characteristic one, e.g. first point, first line, top of flat
roofs etc. The topology of the existing dataset, with the exception of error corrections, should be maintained.
This update should be achieved by using image analysis techniques to be developed at the Institute of Geodesy and
Photogrammetry, ETH Zurich (IGP) and digital aerial imagery. Thereby, IGP makes use of previous experience with
other projects for building and road extraction (Henricsson et al., 1996; Henricsson, 1996; Gruen and Li, 1997), which
however were rather research oriented, not making use of existing information, and regarding buildings were focussing
on large image scales and detailed roof modelling. The whole procedure should be implemented as a stand-alone
software package, able to import and export data as used at L+T. It should be quasi operational, fast, and the most
important reliable. We do not aim at full automation (ca. 8096 completeness is a plausible target), but the "correct"
results should be really correct to avoid checking manually the whole dataset. We are envisaging a traffic light system,
with high reliability for green, red and yellow, and as high as possible percentage for the "green", whereby an operator
should manually check only the yellow, and possibly also red objects. Another important aspect is the use of the current
image data as much as possible (1:30,000 image scale, 15 cm focal length, B/W imagery, 2596 sidelap), to avoid
additional costs. However, it is a topic of the project investigations to find what are the gains with respect to the costs, if
colour is used, 30 cm focal length (i.e. 1:15.000 scale), and 60% sidelap for fourwise image overlap. Use of colour is
not a big issue, but additional flight strips by using 30 cm focal length or 60% sidelap are of more concern.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 463