NG THE
urg,
Tr scanning or
etric images or
he object, it is
stration, which
stages of the
transformation
clouds as well
eatures located
1 this principle
»rrespondences
2 of Simulated
sis confirm the
a sets.
ing additional
son and Kang
ithm by using
o the following
jühring (2002),
neters
the success of
e of automation
) estimates the
er and is not
d b), they are
hotogrammetric
itensity images,
nounted camera
em is calibrated
d, green, blue)
ordinate system
Concerning the combination of the geometric and radiometric
sensors, a new scanning technology is available for several
years. With the availability of color coded point clouds (fig. 2)
or textured objects, the generally higher resolution of the
photogrammetric images offers new possibilities in the discrete
processing stages. In the following chapter a new strategy is
defined in detail, based on experience in digital image
processing and geometric point cloud registration.
Figure 1: Combined sensor — laser scanner with mounted
camera (RIEGL, 2004)
To evaluate the strategy, a data set containing the front of the
main building of the University of Hannover, Germany is used.
The data set is acquired with the Riegel LMS Z360 scanner and
a mounted Nikon D100 with a 6 mega pixel image sensor.
Several view points are available and due to known
transformation parameters, it is also possible to control the
registration step.
NS CH A " NLT
“of i 2. es dr
: e :
4 ?
phe . x me
T. fe vr
HH aes ES niai rb
e 3
Figure 2: Color coded point cloud
2. DEFINITION OF THE MATCHING METHOD
The definition of corresponding candidates is also a question of
the matching method. The task can be separated into low level
and high level strategies. In the low level strategies the
complete original data sets are used for registration. In this case,
it is almost impossible to process the data it in an acceptable
amount of time. In contrast to that, the high level strategies take
much more effort in the preprocessing step to reduce redundant
data and extract the most promising candidates.
In the following, a new operator is outlined in detail, which
combines information from photogrammetric images and 3D
point clouds for registration. It will be shown why the operator
is image based, which role the point cloud plays and how wide
baselines from different view points can be handled by the
operator.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
2.1 Image based point cloud operator (IBPCO)
The first issue is the necessity of invariant features in order to
get the possibility to identify them from different view points.
Much research has been carried out regarding that issue in the
field of computer vision, (e.g. Van Gool et al., 2002, Polleyfeys
et al, 2002, Lowe, 1999). More algorithms have been
developed to extract features in the range images or point
clouds (Lavallee and Szeliski, 1995).
Basically, all these algorithms try to extract distinct edges or
corners in the data sets to identify them from different view
points in a sophisticated manner. The major drawback of these
algorithms is that the occlusion of some features - the
corresponding candidates - causes these algorithms to fail. To
be unable to assess the corresponding candidates before
registration is unsatisfying for automation.
Photogr. À Feature
image 4 extraction
E
A i Calculate 3D
Position
Range Y
image
'——p»4 Plane adjustment
+
» | Create discrete
Ortho-Image
Correspondin
g candidates
r *
Figure 3: Image based search of corresponding candidates
All features
evaluated?
The IBPCO is based on the fact, that high resolution images are
available and oriented in SOCS and the assumption that some
areas exist where sufficient texture for image matching is
