3D City Modelling for Mobile Augmented Reality 
153 
ented Reality 
cenes superimposed 
bjects that is closely 
nd creation of a 3D 
ented reality system 
3D topological data 
groups, i.e. methods 
xuction of buildings 
of buildings, further 
1 manually, utilising 
ages, and automatic 
tically from a large 
suits are deliberated 
ration of data in the 
sing demand for 3D 
-made objects (e.g. 
is suited for all the 
is where 3D models 
tion of a real scene 
nes with additional 
ice known as Head 
if the real world can 
, etc. One of the key 
losely related to the 
;mented reality. The 
iy, The Netherlands 
ich the 3D model is 
ng?”, “where is the 
bone computers for 
. Depending on the 
nting the way to the 
al and real views is 
lal objects that may 
al and real objects. 
del is to be used for 
is based on a vision 
i the inertial system 
'his accuracy is not 
Id like to justify the 
'at least 10 cm. The 
ures extracted from 
JbiCom project, the 
itures for matching, 
le features, i.e. well 
distinguished lines on the facades of buildings, on the ground (e.g. tiles, bicycle paths) or on other objects (e.g. lampposts, traffic 
signs) in the field of view. Furthermore, the number of lines supplied by the 3D model plays a critical role in the matching procedure 
and hence for the accurate positioning. 
Second, the rendering subsystem (for visualising virtual objects) makes use of the 3D model but in a different way. In order to 
achieve realistic visualisation, the virtual objects have to “behave” as real objects, i.e. when occluded by real objects, the 
corresponding parts have to be invisible for the observer (Pasman and Jansen 2001). To achieve this effect, the rendering engine must 
know the exact position of the mobile unit and the position and the shape of the occluding real object. Currently, we concentrate on 
large 3D man-made objects (e.g. buildings, monuments, bridges) as potential occluders. Real objects such as trees, cars, windows, 
doors, balconies, etc. are not considered. The geometric representation of potential occluders has to assure connectivity and 
continuity, i.e. gaps between polygons or polygons with holes are not acceptable since they may disturb the perception of the mixed 
scene. In other words, 3D topological consistency is required. 
The analysis of the role of the 3D model in the augmented reality system can be summarised into three basic requirements as follows: 
• maintenance of topologically structured 3D objects, 
• accuracy in the range of few decimetres, and 
• large amounts of details on facades organised as individual line features. 
In this paper we concentrate mainly on the procedures for reconstructing topologically structured 3D objects with decimetre 
accuracy. 
The test area for the outdoor augment reality system is the central campus area of the Delft University of Technology (see Figure 2). 
Although the area is relatively small, the 3D objects of interest revealed large variations in shape and complexity. Here, we 
concentrate on the reconstruction of five 3D objects (see Figure 2) namely the Aula (1), the Faculty of Applied Physics (2), the 
Faculty of Mechanics (3), the Post-office (4), and the Art monument (5). The terrain objects such tiles, bicycle pats, streets, parking 
lots, gardens, etc. situated in the same area are considered as well. 
Figure 2: Test area: the campus of the Delft University of 
Figure 1: Schema of UbiCom augmented reality system Technology 
The selection of an appropriate approach is dependent on many factors such as required accuracy and detail, complexity of the 
reconstructed houses, availability and accuracy of data (images, maps), time and manpower constraints. Considering the relatively 
high accuracy requirements and the complexity of the buildings in the reconstructed area, we have chosen to apply manual and semi 
automatic methods for reconstructing. Bearing in mind the normal behaviour of a walking person (i.e. looking mostly at objects at 
front and around), we concentrate on facades of the buildings and terrain objects rather than on the roof elements (usually not visible 
from street level). Thus, every building of interest is separately reconstructed as the main attention is on the front facades. In order to 
obtain the position of the individually reconstructed models in the real world and compute the precision of the reconstructed 
buildings, Least Squares Adjustment (LSA) is performed. Moreover, the reconstructed objects are to be organised in a 3D topological 
data structure (Zlatanova 2001). This requires clarifying of spatial relationships during the reconstruction procedure, which in some 
cases led to the development of supplementary software. The initial intentions were to limit the source data to only terrestrial (images 
taken from street level) and aerial images. At later stage, we have incorporated laser altimetry data and the large-scale digital 
topographic map (called GBKN). Thus the reconstruction procedure is a combination of several different approaches depending on 
the type of the objects and complexity of the shapes. Figure 3 gives a general view of the used software and data to obtain the 3D 
topologically structured model. We distinguish between procedures for reconstruction from terrestrial (Section 3) and aerial (Section 
4) images, and procedures for reconstruction from laser data and GBKN (Section 5). The following sections discuss the major steps 
in more details. 
mber 18 - 21,2001