es of the AT
liability and
r factors are
ntion at this
togrammetry
ni-automatic
or IM met all
lis reason, to
ther factors,
f control of
neck on the
ilure instead
-level image
arching and
extraction, a
mages and a
listinguished
of pyramid
e matching,
liability and
ATPS which
With this
rder to meet
1 SIX more
high
duce the
ure 4). At
ssible.
zies to make
roughout the
] by a scale
ly about 3-4,
solution, the
ercentage of
cent images
re to put the
vo/three/five
/hich images
rizontal and
) be used to
essing. The
percentage of the overlap will be updated if the difference
between the initial value and the value from the original
resolution is over a certain threshold.
Searching starts trom the highest level with help of the value
of the initial bi-direction parallaxes, then a selected point
always serves as the approximations of the next level. The
process is subsequently repeated by the remaining levels of
the pyramid images until the matching are successful on all
levels including the level zero (original image), on which the
LSM is carried out to get the final measurement.
Point-like feature extraction and a combined matching
A point-like feature extraction around the selected area of the
central column of each master image (Figure 4) is executed on
the next higher level of the original image, based on which the
locations of the features on other levels are simply derived.
The matching proceeds from the highest level to the lowest
level with a spiral search within a certain range. The spiral
search will be concluded automatically when certain indicators
show no more hope for further improvement. Once the
matching is successful for a feature on all planned levels,
additional feature extraction in a small area around the feature
on the original image, e. g. 5 x 5 of primary image, is carried
out again to get the feature better located. The point-like
interest operator used throughout the system is a simplified
version of the Foerstner's (Lue, 1988)
The reason that the feature extraction is carried out on a
certain higher level rather than on the original 1mage and all
levels of the pyramids is to avoid some possible noise existing
on the original one and also to save effort/time of repeated
feature extraction. With this approach it doesn't need to
concern the feature's accuracy too much because the final LSM
will well take care of it and give even higher accuracy than
any feature extract operator does.
Optional parameters
The entire process of ATPS is automatic and though no
approximations are needed, there are a number of optional
parameters for user to choose; like which pattern (Figure 4),
the size of the patch for feature extraction, the number of the
features to be used, independently matching for each selected
point or based on the info from the existing points. Once the
parameters are defined at the beginning, the tie points will be
automatically selected with a desired pattern and number.
In each defined patch, normally many more features than
needed are extracted in order to avoid loss of any possible
features by automatically setting a relative lower threshold for
the interest value, especially for the areas with few textures
and bad contrast. Therefore the number of tie points for each
patch are highly flexible to be chosen accordingly through
using a proper parameter. On the other hand, an ordering list
of features according to their interest values is made to let the
best features with higher weight value always be treated first.
It should be pointed out that some commercial
photogrammetry companies do not want to have too many tie
points, instead to prefer a minimum but sufficient number of
tie points, especially for the normal areas and applications,
even though the processing is automatically carried out. The
point here is that the quality of the final A7 results are good
enough to meet the requirements, that drove us to provide
different tie point configuration patterns for different users to
choose. Total seven patterns for tie points distributed in a
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
homologous image segments are provided to give the user a
high degree of flexibility (Figure 4). This is also useful when
some areas fall into the water or other difficult fields like
forest, brushes, shadows etc. Points selected from the
additional areas will compensate the lost points in other areas.
Quality control
The algorithms set up several checks on their own, such as the
history of the matching for each pyramid level, LSM's accuracy
indices etc. An affine transformation is also applied to identify
the possible mismatches after enough points are selected.
It should be pointed out that no ground control points (GCPs)
or any point approximations are needed for ATPS, though a set
of transformation coefficients can be derived from existing
GCPs to partly reduce the effort of tie points searching.
Similarly, GPS data, if available, would also play an important
role to make the searching more effective.
Interactive tool
Unlike AIO, ATPS can not easily reach hundred percent
success rate all the time, a certain interactive involvement by
the human being is often needed, especially in the areas of
forest, urban area with a large image scale that normally
causes a large perspective distortion. In order to compensate
the loss of tie points in such areas SoftPlotterTM provides
users with a very convenient tool to display three (for a single
strip) or six frames (for two strips) altogether on a computer
monitor to let user check the quality or add/delete any points
easily.
The selection of tie point with this tool is a so-called semi-
automatic one. Its automatic level depends on how many
points (tie points or ground control points) are already exist.
Normally, it is recommended to run ATPS first, because many
points selected by ATPS will be used to calculate the proper
transformation parameters to issue a prediction for any desired
position. For instance, to add one point that is required by an
user is to select a point manually only from the master image
and then click a special button (“Auto Place” or "Pug Point")
that triggers a series of calculations, including roughly locating
all corresponding points from the homologous images and
conducting the LSM. Then, all conjugate points will be
automatically and precisely found from the slave images and
simultaneously displayed on the screen.
2.2 Results
Total of 16 data sets with 60-80 percent overlap were tested
with ATPS so far. All data sets were b/w, with the exception
of one color data set. The image scale covers 1:3400, 4000,
4300, 6000, 9600, 24000, 45000 which were scanned by
different scanners with different formats and resolutions (15,
22.5, 25, 30 microns per pixel) The data sets containe all
kinds of different textures, including urban area with strong
distortion, forest, mountain, wetland, desert area, rivers, lakes
etc.
The average success rate is about 85 percent. Some data sets
even achieved hundred percent success. The average
processing time was less than a half minute per frame with
points on nine standard positions on SGI Indigo2 with
R4400/200 MHz worksattion.
