International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012 
XXII ISPRS Congress, 25 August - 01 September 2012, Melbourne, Australia 
117 
VHRS STEREO IMAGES FOR 3D MODELLING OF BUILDINGS 
A. Bujakiewicz *, M. Hole 
Department of Photogrammetry, Remote Sensing and Spatial Information Systems, Faculty of Geodesy and Cartography 
Warsaw University of Technology, Plac Politechniki 1, Warsaw, Poland 
abujak7@wp.pl, malgota@wp.pl 
Commission IV/3 
KEY WORDS: Orientation, Extraction, Modelling, High Resolution, Ikonos 
ABSTRACT: 
The paper presents the project which was carried out in the Photogrammctric Laboratory of Warsaw University of Technology. The 
experiment is concerned with the extraction of 3D vector data for buildings creation from 3D photogrammctric model based on the 
Ikonos stereo images. The model was reconstructed with photogrammctric workstation - Summit Evolution combined with ArcGIS 
3D platform. Accuracy of 3D model was significantly improved by use for orientation of pair of satellite images the stereo measured 
tie points distributed uniformly around the model area in addition to 5 control points. The RMS for model reconstructed on base of 
the RPC coefficients only were 16,6 m, 2,7 m and 47,4 m, for X, Y and Z coordinates, respectively. By addition of 5 control points 
the RMS were improved to 0,7 m, 0,7 m 1,0 m, where the best results were achieved when RMS were estimated from deviations in 
17 check points (with 5 control pointsjand amounted to 0,4 m, 0,5 m and 0,6 m, for X, Y, and Z respectively. The extracted 3D 
vector data for buildings were integrated with 2D data of the ground footprints and afterwards they were used for 3D modelling of 
buildings in Google SkctchUp software. The final results were compared with the reference data obtained from other sources. It was 
found that the shape of buildings (in concern to the number of details) had been reconstructed on level of LoDl, when the accuracy 
of these models corresponded to the level of LoD2. 
1. INTRODUCTION 
In the last decade the Very High Resolution Satellite images 
have proved to be very useful for generation of cartographic 
products. The popular satellite orthomaps are usually used for 
mono plotting of 2D vector data. The accuracy of such data is 
compatible with the accuracy of middle scale aerial 
photography, however their content is reduced and the object 
elevations can not be determined. The procedure of image 
orientation can be performed in various environments, for 
example PCI Geomatica. This software applies the RPC 
coefficients together with a number of control points measured 
on each image separately. Development of stereo digitalization 
modules of photogrammctric digital workstations has made 
possible to execute the stereoscopic measurement of data and 
therefore to increase orientation accuracy of VHRS stereo pair 
images. Consequently it is possible to acquire 3D products, such 
as DSM/ DTM and vector data for buildings reconstruction and 
modelling (Eisenbeiss and others, 2004; Toutin,2004a,b; 
Kaczynski and others, 2004; Zhang, Fraser, 2008). In general, 
according to (Kolbe and others, 2005), 3D models of buildings 
can be reconstructed in respect to five levels of details (LoD) 
defined in CityGML scale, due to their accuracy and details’ 
description. In research published by (Ulm, 2003; Kocaman 
and others, 2006), it has been found that 3D building models 
from stereo satellite VHRS imagery with 0.5 - 1.0 meter 
resolution, such as QuickBird or Ikonos-2, can be reconstructed 
with standards of LoDl - LoD2. In case of buildings 
reconstruction on level LoDl, a simple shape of block models 
with flat roofs and the accuracy of 5 meter in horizontal (XY) 
and vertical (Z) positions, are required. For level of LoD2, the 
texture model with real shape of roofs and accuracy of 2 meters 
for XY and 1 meter for Z, arc requested. 
In research carried out by (Kocaman and others, 2006), the 
specific difficulty for recognition of roofs types and their small 
features, was found. The example of area with 3D building 
models reconstructed from the QuickBird images with the CC- 
Modeler is shown in figure 1. 
Fig. 1 3D city model from QuickBird images reconstructed by 
the CC-Modeler (Kocaman and others, 2006) 
2. PRINCIPLES OF THE APPLIED APPROACH 
Reconstruction of 3D models from VHRS satellite images 
requires their exterior orientation, similarly as in classical aerial 
compilation. In case of aerial imagery, the exterior orientation 
parameters for a block of photographs, which cover the area, are 
usually determined by aerial triangulation with support of GPS 
or GPS/INS flight data. In case of VHRS imagery, the methods 
of sat-triangulation have been also developed (Dial, Grodecki, 
2004, 2004; Ewiak, 2009). However, the area covered by a 
satellite image is much larger and very often only single images 
are sufficient for the area coverage. In such cases, a block 
triangulation approach is not required but simpler methods of 
exterior orientation determination can be used. The RPC 
coefficients, delivered with the satellite images, serve only as 
the approximate data and do not supply the sufficient 
orientation accuracy. In recent years many research works have 
been executed on determination of exterior parameters for 
single images with use of a typical remote sensing software (for 
ex. PCI Geomatica). They have usually required a well