International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
available. The goal is to identify reliably corresponding
candidates by using the radiometric and geometric observations
of the data set. The algorithm is based on the assumption that
only features extracted in photogrammetric images and which
are located in planar geometric areas could be suitable
candidates (cf. fig. 3).
In the first step, distinct features have to be extracted with an
appropriate operator for digital image processing. In this work
the SUSAN (Smallest Univalue Segment Assimilating Nucleus)
operator developed by Smith and Brady (1997) is used. In
principle all pixels within a circular mask are compared with the
nucleus. Therefore a threshold has to be set according to the
contrast and noise of the intensities to assign the pixels to the
nucleus. The sum n of the comparison will be compared with a
second threshold g. For a corner to be present, n have to be less
than half of its maximum. Shortly, g predicates the geometric
quality and / the density of the features.
Then the 3D position of the feature is interpolated from
surrounding points (cf. fig. 4). For efficient processing the
points are transformed into the image space of the camera under
consideration of lens distortions of the used camera. As well as
the position, an adjusted plane is estimated from these points.
The following inspection of visibility and smoothness is used to
test the suitability of candidates (fig. 5). Depending on the set
threshold for the max. viewing angle, the max. enabled baseline
can be derived from. Features located in geometrically
discontinuous areas can be separated due to points which differ
from the estimated plane. These points can be detected with
blunder detection in the plane adjustment or by calculating the
curvature, the cosine angle between the normal vector and the
difference vector of the interpolated 3D position of the SUSAN
feature and the included points.
Figure 4: SUSAN features surrounded by 3D points. It shows
some possible candidates with there surrounding points.
For the matching of the features which fulfill these conditions
later on, a discrete orthoimage is calculated, consider the object
space as is typically done in image matching, (e.g. Heipke,
1992). With this discrete image patch with an adequate size a
reliable correspondence check is possible.
Therefore, firstly a grid is defined on the plane in object space,
whereas the metric size of one facet is depending on the set
image scale. Secondly, to fill the grid, each facet will be
mapped into the considering image, then the grey values will be
resampled bilinearly.
\
\3
V.
N 4A
=
3D point
e rez
m— — M — — SE —
plane — —@ @ ——6
BT 2 M
-—7— M — —R— € nn“
e
Figure 5: Visibility and Smoothness. The rulings of the
geometric acceptance are defined on the one hand by the angle
resulting from the scalar product between the normal and the
image ray, and on the other hand by the max. absolute residual.
Here, also an inspection is necessary. Since the unavoidable
comparability of the discrete orthoimages with candidates from
different view points a fix image scale is set. The generation of
the orthoimage can also fail if the SUSAN feature is located at
the border of the image, meaning that not enough information is
available to fill the grid.
However, until all features are evaluated the matching can be
executed.
3. EXECUTION OF THE MATCHING METHOD
To match two sets of features from different view points,
reliable criteria have to be defined to find the correspondences.
Therefore a couple of possible criteria have been developed for
this process based on the point clouds and the photogrammetric
images.
3.1 Assessing criteria
Radiometric uniformity:
Is calculated with the cross correlation between the
corresponding candidates. Due to the fact, that the patches have
the same geometric resolution and the same orientation in
object space, only the brightness have to consider what is
regarded by the cross correlation. The resulting coefficient
assesses the uniformity.
Oo
=, Ig 1)
Pg
0,0,
Intensity:
The scanning device supplies the intensity for observed 3D
points, with which the active signal is reflected on the surface of
the object.
E, 7 Abs(C, — P.) (2)
Neighborhood:
The assessment of the neighborhood is only possible in a final
post process. It can be done directly in three-dimensional object
space or more efficiently in image space. Due to the rectified Z-
International Archiv
UE TE
axis (perpendiculari
the different view pc
corresponding candi
is identical, which w
Erop = J (Pueighbor
3.2 The matching
Due to the fact tha
controlled by the t:
strategy for the m
combinations. This
whose influence is ri
For combinatorial cr
(Huber et al. 2001)
Wendt, 2003, Luck
optimization metho
RANSAC (Random
select the minimum
parameters and asse
when the consensus
In the following the
to give an idea about
3.2.1 Simulated A
The term annealing
cooling of material, «
cooling leads to a re
while a fast cooling
level. In a controll
energy of the mater
reduced to a global r
transfer of this prin
corresponding candi
6.
The solution for tl
initialized by randon
initial solution a s
temperature 7; is det
has been derived in
correspondence wil
, temperature 7, accol
approximately 0.5 ev
D^ AE*LA
Starting from this ten
has reached the equi
does not decrease
iterations have been
new corresponding
give rise to a change
solution. A new so
decreases. Otherwise
TAE y,
W(AE,,. ) zi ai