on detection and selection of signals and which addresses
identification, precise measurement and verification is still
far from being solved. For more details please confer Hahn
(1993) and Gülch (1995).
In the concept presented in the next section we assume
that a human operator is involved in the measurement of
signalized points.
2. A CONCEPT FOR SEMI-AUTOMATIC
MEASUREMENT OF SIGNALIZED POINTS
Depending on the block configuration and the position
within a block signalized ground control points, check points
or tie points are imaged, for example, in two, three or six
photographs. Only the points in the extreme corners of the
block may appear in only one image. Therefore, we aim at
a procedure for multi-image measurement of the signals in
all images.
The proposed concept
1. collects templates which represent the various types
of signals in a library,
2. supposes that the approximate position of the signal
in one image is given, e.g. measured manually by a
human operator,
3. uses area based multi-image matching to provide the
approximate location of a signal in all overlapping im-
ages and
4. measures the precise image location of a signal by
matching with templates of the library.
In Digital Photogrammetric Systems (DPS) today it is quite
standard that image pyramids for all images of a block are
available. Further parameters which characterize the block
structure completely (like image and strip parameters, flight
direction, neighbourhood of images and strips, overlap in p
and q, etc.) are also represented in the system. With the
image pyramids it is guaranteed that matching processes
can operate on different resolution levels. Together with the
knowledge on the block parameters coarse-to-fine match-
ing allows automatic point transfer in all overlapping im-
ages. Tsingas (1991) has shown that this works well in a
feature based solution. In cooperation with the Zeiss com-
pany we developed an area based least squares solution
for stereo and multi-image matching. With this algorithm
matching of one point pair through 7 levels of an image
pyramid is performed in about 1.2 seconds on a Silicon
Graphics Indigo2 Workstation (R4400/200MHz).
The measurement of a signalized point is started with a
semi-automatic preparation step in which the approximate
location in one image is defined by the human operator.
The determination of the location in all overlapping neigh-
bouring images is solved with the multi-level least squares
approach mentioned above.
The measurement process for localization of the signalized
points consist of a two step solution. In the first step area
based matching (with 6 geometric and 2 radiometric pa-
rameters) between the template T and the images 1 .. N is
carried out (figure 1).
292
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Figure 1: Step 1: template matching
With the estimated transformation parameters the refer-
ence point of the template is transferred into all matched
images. Then matching back is carried out by changing the
role of mask and search image. Because the least squares
solution is solved by linearization and adjustment over the
estimates in the search image this small unsymmetry may
lead to differences in the estimated results of both solu-
tions. In general, the differences will be small because a
prerequisite for matching back is that matching from mask
to search was successful. If nevertheless a certain differ-
ence is obtained this indicates that differences in the image
structure of mask and search have an influence on the esti-
mated optimal solutions through the gradient spaces (which
are the spaces in which the optimal solution is 0). In addi-
tion to the standard convergence criteria inconsistency with
matching transferred points back is taken as internal mea-
sure for assessing successful matches.
If matching of the template with an image fails this may lead
to a situation like that sketched in figure 2.
Eoi ced
4
x .
|
i
Figure 2: Step 2: completion of failed matches
In this case the pairs T-1, T-3 and T-N are processed suc-
cessful whereas matching T-2 and T-4 has failed. Now the
second step of the procedure is activated which aims at
completing this failures. For that, all successfully matched
points of step 1 are now used to carry out image to image
matching. At the example of figure 2 this is matching of
the pairs 1—2, 3-2 and N-2. If this succeeds for more than
one pair then the mean of all estimated locations in figure
2 is calculated. In the same way the process is continued
until all missing correspondences are supplemented or all
possibilities are exhausted.
The role of the template library is not addressed so far.
From a practical point of view it is quite simple to take a
given library into account by substituting the template con-
sidered so far with a series of templates. But this leads
to an interpretation problem because associated with each
templat
able mc
ing with
match
the loca
For the
module
- with 0
fect,
- with 1
- with 2
way (st
- With 1
The sul
templat
as the
a speci
descrip
ments.
In prac
for an
Scanni
e.g. wi
aged te
targets
O
Figure
shows
21 pix
Furthe
