tie point extraction process). A high level of automation can be
achieved.
In the system at FGI the tie point areas are determined in two
steps. First the overlap areas are coarsely defined by using
approximate coordinates of the perspective centres. The
coordinates of the tie point areas are usually not accurate
enough for tie point extraction, because the approximations may
be rough and there are normally variations in the elevations of
the terrain. The coordinates are therefore refined by image
matching. In the refinement, cross correlation with a special
matching strategy on low resolution images is used. Possible
gross errors are detected in the block adjustment and additional
observations are carried out interactively. The approach has
been proven to work well, for instance in the OEEPE test block
Forssa, only 2.2 % of the automatically measured observations
were erroneous, see (Honkavaara and Hggholen 1995).
2.1.2 Corresponding point definition
The tie points are measured in tie point areas using image
matching. There are different matching methods available, from
techniques using local image information, like least squares
matching (LSM) and feature based matching (FBM), to global
techniques, see overview in (Forstner 1995).
At FGI, the approach selected for tie point extraction is based
on the one developed by Tsingas, see (Tsingas 1992, 1994). The
tie points are first extracted using multiple image FBM. Because
of the rather low accuracy of FBM, the extracted coordinates
are refined by LSM. The Tsingas’ method is further refined to
achieve higher speed and good success rate with multiple
matches, see (Honkavaara and Hggholen 1995). In order to get
good enough approximate values for the matching process a
multiresolution image pyramid with 3 layers (scales 1:16, 1:4
and 1:1) is used.
2.1.3 Block adjustment and point selection
The block adjustment can be seen as an important part of the tie
point measurement process. In the block adjustment, in addition
to solving the unknowns of the mathematical model, also the
possible gross errors are detected.
Automating the block adjustment when using interactive
measurements is treated in (Sarjakoski 1988). Automatic tie
point measurement gives new features to the block adjustment
like:
1. there are considerably more observations,
2. the quality of the observations is unknown and
3. there are more gross errors.
It is important to investigate if the techniques developed for
interactive measurement are sufficient when using automatically
measured observations. The experience gained so far from block
adjustments with automatic tie point observations is that in
some cases the use of additional parameters have negative
influence on the accuracy of the block, see (Honkavaara and
Hggholen 1995). Additional parameters are sensitive to
inaccurate observations and poor distribution of the tie points.
Tests have shown that correct weighting is critical as is the type
of additional paranieters to be used.
In the system at FGI the whole block is processed in a single
block adjustment using a separate adjustment program. The
adjustment procedure does not differ from the one used with
338
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
interactive measurements. Block adjustment will be
implemented as a part of the tie point extraction program. This
means that block adjustment will be performed also on sub
blocks, which makes the quality control easier.
More observations are measured than needed in the tie point
extraction process. Consequently, an important task after the
block adjustment is to select a sufficient number of relevant
observations. At the moment the selection process relies on
heuristic ideas. The criteria in the selection are: 1) importance
(number of observed images with insufficient number of
observations), 2) completeness (maximising the number of
observed images), 3) distance from other selected points (good
distribution) and 4) distinctness (a minimum requirement is that
the LSM windows do not cover each other).
2.1.4 Quality control
In the traditional interactive aerial triangulation process, the
operator ensures that there exists a sufficient number of good
observations. The final quality control is performed in the block
adjustment.
In the automatic measurement process, in addition to checking
the statistics, also the adequacy of the observations in each tie
point area have to be checked, i. e.:
1. The number of observations in each image is sufficient. Tt
has to be checked, that there are enough observations in
each of the overlapping images.
2. The completeness of the observations is sufficient. It is not
enough to check only the number of observations in each of
the images. As mentioned in Section 2.1.1, to achieve
stability in the block, also matches on multiple images are
needed.
3. The distribution of the observations is sufficient. The tie
point observations should have proper distribution.
Sufficiency, completeness and distribution of the tie point
observations are under investigation at FGI. They are discussed
in Section 2.2 and empirical results are presented in Section 3.2.
2.1.5 Processing of the unsuccessful tie point areas
The tie point areas which failed in the quality control have to be
treated. In general, the matching may fail because of difficult
objects (difficult 3D-object, monotone object, water, forest,
obstacle etc.) or poor imagery (radiometric differences etc.).
The matching method affects the rate of failures.
Different actions can be carried out in the failed areas,
depending on the reason for failure:
l. Regard the failure as non-influent. The failure does not
deteriorate the accuracy.
2. Search for better location for matching. Regardless of
failed matches in certain locations, there exists good areas
for matching for the given image combination in the overlap
area.
3. Search for optimal image combinations. There exist no
match for the given combination of images in the overlap
area, but there exist matches for some other combinations.
4. Stop matching with impossible images. Matching will not
succeed in the overlap area in some, or in the worst case in
all of the images.
5. Select another matching method. Matching may succeed
using another method, for instance, by interactive
measurement.
To perfo
system. *
In the sy
The rea
mention«
2.1.6 Pr
Knowlec
realised
this kno
extractio
At the s
moment
I.
2.
Defi
Sect
Extr
see!
Perf
poin
Che
Proc
Itera
with
2.2 Abo
point ok
Importa
distribut
observat
imagery
accuracy
are brief
2.2.1 Di
The con
treated «
distribu!
means t]
tie point
When
measure
position
conside;
automat
hand, tk
using al
more d
increase
images
222N
The nui
extracti
measure
observa
concern
achieve
accurac
poor (F
