CI PA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey
Figure 6: Smoothed surface with automatic generated
NURBS patches. The black lines show the patch bound
aries.
point phase. It seems to be reasonable to invest that time
into point phase because triangulation needs a well-edited
point cloud, otherwise it is hardly possible to overcome
the difficulties arising at polygon phase.
5 DISCUSSION AND SUMMARY
With the help of hybrid adjustment techniques a si
multaneous transformation of the image and laser
scanner data into a global co-ordinate system on the
basis of homologous points is possible (cf. Sec. 3).
Additionally, we used variance component analysis
for the determination of the individual accuracies of
the different observation types within the adjustment
procedure. These accuracies can be compared with the
nominal accuracy of the measurement sensors (e.g.: The
computed variance of the range measurements to the
signalized tie points of the fine scans is the same than the
nominal accuracy of the laser scanner Riegl LMS-Z360.).
Furthermore, the analysis shows that the accuracy of the
coarse panorama scans is lower than the one of the fine
scans and that the manual measurements in the digital
images of the signalized tie points have sub-pixel accuracy.
The generation of the object model of the statue
Marc Anton was performed with the software package
GeomagicStudio on the basis of the laser scanner point
cloud. During the modelling process we have recognized
that the laser scanner has obviously problems with
measurements close to the contour of the objects (cf. Fig.
4). This can be explained by the beam diameter and the
fact, that only a part of the energy is reflected from the
object surface, the other part is lost, or reflected further
behind. As the intersection angle between the surface
and the laser beam comes closer and closer to 180 °, the
reflection area on the surface becomes bigger, too. This
may also explain the deterioration of the measurement
quality.
In general the triangulation of the data did work
quite well, but there occurred problems in occlusion
areas. In these regions time consuming manual editing
was necessary. Additionally problems were caused by
random measurement errors (approx. ±2cm dispersion of
the point cloud around the averaging surface). Due to
this noise local curvature based re-sampling methods did
fail. Therefore smoothing operations were necessary. As
a result, the final surface model is a compromise between
smoothness and detail preservation. The difference
Figure 7: Color coded difference model between the final
triangulated surface and the original laser point cloud data
(units: m).
between the final triangulated surface and the original
laser scanner point cloud can be seen in the Fig. 7. The
average distance of the points to the model is 12mm
and the standard deviation of these differences is 13mm.
Finally, a NURBS surface is automatically computed on
the basis of the triangulation. This final CAD-model (cf.
Fig. 6) describes the statue by a continuous surface (in
the case of our statue it has holes due to missing data).
This model is differentiable within the boundaries of one
NURBS-patch, whereas on the patch boundaries this
differentiability is only approximatively given (epsilon
tolerance).
On the basis of our experiences gained during the
modelling step, we can say that it would be very useful to
integrate additional image measurements into the mod
elling process. Especially the accurate representation of
edges on the statue would be much easier with the help of
additional image data. As mentioned above, the distance
measuring unit of the laser scanner has difficulties with
measurements close to the contour of the object. It would
be possible to eliminate this lack by the digitalisation
of these line features in the image data. This capability
is demonstrated in Fig. 8. For this visualization the
contours of two lions were digitized in one image. Then
for the determination of the 3D co-ordinates these contour
points are projected into a vertical plane. Finally, these
points are connected to the projection center of the digital
image. This leads to a general cone which is displayed
with the help of a VRML-visualization. Additionally
to the possibility of accurate 3D point or line/curve
measurements the image data provides texture informa
tion, which can be very important for interpretation and
documentation. Therefore the next aim for the docu
mentation of such an object will be the task to project
the texture information onto the final surface model in
order to generate a photo-realistic 3D model of the statue.
Summing up our experiences in the combined use of
laserscanning and photogrammetry for heritage recording
and documentation it turned out that both techniques
are necessary for the determination of a high quality 3D
model. With the help of hybrid adjustment techniques all
observations can be simultaneously transformed into one
global co-ordinate system ([ORPHEUS/Orient, 2003]).
As mentioned at the beginning, the modelling of the
statue Marc Anton is - due to its complexity - a very
difficult task. During data capturing we had to accept
data holes in occlusion areas to reduce the time for
