(4.13) 
problem of 
ining main 
in different 
1 processing 
; and control 
| solution of 
It is proved 
ss the series 
use of GPS 
e process of 
(3.1) takes 
'rammetrical 
cases of this 
s of exterior 
ns, software, 
imocreñ. M 
pus.- 1982 : 
X.//Teone3M, 
//- Teone3m, 
13HpOBAHHB! 
HayK. JIbBok 
DO pAHHaTHb) 
xondepenll 
177-189. 
KapTorpadif 
Markus Niederöst 
RELIABLE RECONSTRUCTION OF BUILDINGS FOR DIGITAL MAP REVISION 
Markus NIEDERÖST 
ETH Zürich, Switzerland 
Institute of Geodesy and Photogrammetry 
markus € geod.baug.ethz.ch 
Working Group III/4 
KEYWORDS: Buildings, Blob detection, Classification, Vector data, Reconstruction, Map revision 
ABSTRACT 
This paper describes a framework for automatic reconstruction of buildings in order to correct and update an initial 2-D 
vector data set and to derive a 3-D model for visualization. The color images of a stereo model are used to calculate a 
digital surface model and an orthophoto. A procedure for blob detection from the calculated digital surface model as well 
as unsupervised multichannel classification are used to produce additional approximate vector data. This allows to 
introduce buildings to the data set that were not yet included. Methods of digital image processing are used to improve 
planar location, rotation and scales of the vector data for each single building. The solutions are rated in order to keep the 
good results and to reject inaccurate or wrong data automatically. Finally one representative height for each building is 
derived from the digital surface model. The accuracy of the building reconstruction is determined by quantitative 
comparison with reference data, and the rating procedure is checked. The described system neither requires large image 
scales nor threefold up to sixfold overlap nor laser scanner data nor infrared information to automatically revise the 
initial vector data and to produce a coarse 3-D city model. Approximation vector data can be used, but it is not 
mandatory. 
1 INTRODUCTION 
The work presented in this paper is part of the project ATOMI (Automated reconstruction of Topographic Objects from 
aerial images using vectorized Map Information), a cooperation between the Swiss Federal Institute of Topography, Bern 
(L+T) and the Institute of Geodesy and Photogrammetry (IGP) at ETH Zürich. Primary aim of the project is to contribute 
to the production of a 3-D landscape model of Switzerland including buildings and roads. Input data to be used are aerial 
images which are produced for the whole area of Switzerland in a 6 years cycle. Approximation vector data is acquired 
by L+T from digital raster maps (scale 1:25'000) by means of semi automatic vectorization procedures. The questions 
that are being investigated at IGP are - for roads and buildings - to automate the refinement and revision of the vector 
data. Besides the improvement of accuracy this also includes the determination of one representative building height and 
the completion respectively update of the vector data set. For a general overview of the project see (Eidenbenz et al, 
2000), the part about roads is described in (Zhang/Baltsavias, 2000). 
Reconstruction of buildings is currently one of the main topics of the photogrammetric community, and the approaches 
to solve the problem as well as the used kind of data are manifold. (Henricsson, 1996) used fourfold overlapping colour 
images (pixelsize in object space 7.5 cm). A procedure with even sixfold overlap is described in (Baillard et al, 1999) 
where the ground resolution is 8.5 cm. Both approaches take profit of the high ground resolution and the number of 
images and thus reduce the problem of edges that are not contained in some of the images. (Weidner, 1997) or (Maas, 
1999) use very precise digital surface models from airborne laser scanning systems, while the classification procedures 
of (Haala/Walter, 1999) base on color infrared images which simplify the separation of man-made objects from 
vegetation. An overview of various approaches can be found in (Griin, 1995) and (Grün, 1997). 
In the project ATOMI neither large scale images with high resolution or infrared images are used nor is the digital surface 
model produced by laser scanner or radar. Due to economical reasons the procedures being developped have to reliably 
work in regions with twofold overlap. On the other hand a priori knowledge (approximate vector data) can be used. Aim 
of the work described in this paper is not to get a fully automatic system but a very reliable system. Buildings that cannot 
be automatically reconstructed with sufficient precision have to be automatically rejected and can be remeasured 
manually. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 635