'anbul 2004
ped to the
ng, a suffi-
ed over the
d in the im-
AN camera
e sub-pixel
ments only
ur opinion,
and image
> geometric
el of detail
els used for
model pro-
"the visible
for the fa-
ical for 3D
1e data.
image.
ality of the
ction for a
«ture from
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Figure 5 exemplarily demonstrates the visual quality of virtual
images generated from the available building models, which are
textured using the collected panoramic scene. In order to gener-
ate the virtual image a standard tool for 3D visualisation was
applied. For this purpose correspondences between image and
object points were provided after the visibility of each building
polygon was controlled based on an occlusion detection and
analysis of normal vector of each building polygon.
Usually, these standard tools for visualisation do not allow for
projective transformation during texture mapping. Thus, per-
spective distortions are frequently eliminated in advance by a
rectification of the respective image sections. For our scenes
these distortions could be neglected during the visualisation
process, since the facades of the buildings only cover small sec-
tions of the applied panoramic images.
3. MODEL BASED GEOREFERENING AND IMAGE
SELECTION
As it was demonstrated in the previous section, panoramic im-
ages can provide texture for a considerable number of buildings
from a single camera station. If this camera station can be se-
lected carefully, the quality of image texture is at least sufficient
for visualisations of buildings in the background of the virtual
scenes. Nevertheless, the realisation of closer virtual viewpoints
will require the integration of additional images in order to
minimize disturbing effects i.e. due to occluding objects or un-
favourable perspectives. In order to allow for an efficient proc-
essing of these image, which were collected from standard ter-
restrial cameras, additional tools were developed to support the
geoereferencing and image selection process.
3.1 Model based refinement of direct georeferencing
One important prerequisite for texture mapping is the availabil-
ity of image position and orientation at a sufficient accuracy.
While in our experiments the exterior orientation of the pano-
ramic scenes is determined by spatial resection using manually
measured control points, direct georeferencing can be applied
alternatively if suitable sensors are available. In principle, the
collection of georeferenced terrestrial images in urban areas is
feasible at high accuracies by integrated DGPS/INS systems.
This type of sensor system is for example applied by (Bosse et
al 2000) for the collection of images, which were subsequently
used in order to collect the geometry and texture of the depicted
buildings. Alternatively to the application of such precise but
expensive sensor systems, the accuracy requirements for direct
georeferencing can be reduced considerably if — as in our case -
a 3D model of the buildings at the site is already available. In
that case, the approximately measured exterior orientation can
be refined by an automatic alignment of the terrestrial image to
the depicted building models.
In order to provide an approximate exterior orientation from di-
rect georeferencing, a low-cost system based on a GPS receiver,
an electronic compass and a tilt sensor is applied. Despite the
good theoretical accuracy of differential GPS, there are a num-
ber of practical limitations of this techniques within built-up ar-
eas. Shadowing from high buildings can result in poor satellite
configurations, in the worst case the signal is lost completely.
Additionally, signal reflections from close buildings result in
multipath effects, which are further reducing the accuracy of
GPS measurement. In our experiments position accuracies of 7-
10 m were achieved. The applied digital compass is specified to
provide the azimuth with a standard deviation of 0.6° to 1.5°.
However, compasses are vulnerable to distortion, because espe-
cially in build-up areas the Earth’s magnetic field can be influ-
enced by cars or electrical installations. These disturbance can
reduce the accuracy of digital compasses to approximately 6°
(Hoff & Azuma 2000).
Figure 6: Coarse model to image mapping based on direct
geoferencing
Figure 7: Building localization based on shape based matching.
Figure 6 shows an overlay of the available building model to
the captured image based on the measured orientation parame-
ters. The limited mapping quality, which results from the re-
stricted accuracy of the directly measured exterior orientation is
clearly visible. This coarse model to image mapping can be im-
proved by a Generalized Hough Transform (GHT) was applied
(Haala & Bohm 2003). For this purpose, the coarse model to
image mapping is used to extract the visible silhouette of the
depicted building. The outline of the projected building model
then can then be localised based on the GHT, which automati-
cally provides the relevant two-dimensional transformation pa-
rameters for shift, rotation and scale in relation to the respective
image. The result of this step is depicted in Figure 7.
The GHT additionally allows for a certain tolerance in shape
deviation. This is useful since the CAD model of the building
provides only a coarse generalization of its actual shape as it is
appearing in the image. After the localization of the outline of
the building model, check points can be generated automatically
based on the 3D coordinates of the visible building and used for
