the algorithm tries to locally detect areas where a large
number of points can be matched.
Basically, the accuracy and the reliability of the block
adjustment depends on the block geometry, the number
and distribution of the transferred tie points and their
associated measurement precision (see details
Ackermann, 1996b). Typical numbers for our approach
are 100 - 200 or even more points per image. In order to
keep the number of matched and transferred points at a
certain level, the system automatically creates new tie
point centers in the case of too few points being initially
matched. Thus, the approach is focused in first instance
on the feature-based matching technique. The simple
key idea behind is to attain highly accurate results in the
block adjustment by high redundancy. However, some
circumstances like poor texture, for instance, might
cause a drastical decrease of the number of matched
points. In such cases the least squares matching method
appears as a remedy to obtain a maximum measuring
precision.
The matched points do not have the same meaning as
the points which are usually measured by an operator.
The automatically measured point is located, for
instance, on house corners or in the neighborhood of
natural features (Figure 6 and 7). As long as the result of
the automatic aerial triangulation is used within a digital
system, the adjusted points are not of interest any more.
However, they are important if the digital aerial
triangulation is to be used with analytical plotters.
3. Controlled tests
3.1 Initialization
The initialization of the MATCH-AT system based on an
integrated DEM generation was tested with a small block
of 16 images arranged in 4 strips with 6096 side lap and
6096 end lap. The photographs scanned with 30 pim pixel
size had a photo scale of 1:11500. The block area of
approximately 3.2 km by 3.2 km was slightly hilly with
height differences of 150 m.
In order to simulate a flight index map we derived the
exposure centers from a given aerial triangulation and
randomly varied them with a standard deviation of 100
m. A first crude initialization of the tie point areas was
then derived solely from the strip azimuths, the
approximate exposure centers and an average terrain
height. Thus, this "zero" initialization represents the
coarsest alternative being possible in the concept of
chapter 2.3. After this first initialization some of the tie
point areas were significantly shifted against their
homologous position by more than 1 cm (Figure 8). The
reasons were mainly camera attitude angles of more
than 4 degrees and in second instance slight height
undulations. If one would apply the kernel system with
those initial tie point areas, it could happen in some
areas that only few or even no points would be matched.
Thus, we applied the initialization of the MATCH-AT
system starting at a pixel resolution of 960 um. The
system applied the matching scheme of the kernel
system in combination with a subsequent DEM
generation at the pyramid levels of 960 um and 480 pm.
After each DEM generation new tie point areas were
409
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
created using the updated image orientations and the
terrain surface. At the 960 um pixel level the system
iterated the algorithm twice to obtain a sufficient
convergence of the tie point locations. Table 1 shows the
results of the two pixel resolution levels.
Pixel size op (Block) ODEM
960 um 230 um = 2.6 m 3.4m (1.3 00)
480 um 113 um = 1.3m 1.7 m (1.3 ox)
Table 1: Initialization with block adjustment and DEM
generation
The sigma naught of the block adjustment (7 c» (Block)),
which represents the precision of the automatic point
identification, was in both pyramid levels approximately
0.24 pixel. Those values are equivalent to 230 um and
113 um, resp. The system reported also a theoretical
accuracy of the block DEM posts (=opem) of 3.4 m and
1.7 m. Those values are - as expected - by the factor 1.3
larger than the sigma naught of the block adjustment.
After this initialization the new tie point areas covered
excellently homologous image patches (Figure 9). The
locations of those tie point areas were accurate enough
for a successful image matching in the kernel system,
whose result of the block adjustment is not reported
here. Although those preliminary results of the
initialization part of MATCH-AT refer to a block with a
side lap of 60 96, the method seems to be suited to cope
with standard side laps of 2096 or 30 96, even if the photo
scale is quite large and height undulations cause relief
displacements. The critical case is given if only a coarse
flight index map is provided with large-scale
photography. However, this is going to be more and
more a minor issue because of the GPS technology.
Altogether, these first results are very promising and
future work will be focused on optimizing this integrated
DEM approach towards an efficient and flexible
initialization of the kernel system.
3.2 Block adjustment results
The MATCH-AT system was applied to two blocks
comprising 43 and 21 images, resp. The overall goal of
the controlled tests was in first instance to assess the
accuracy potential of the automatic aerial triangulation
under practical conditions, especially with respect to the
number of block images to be processed. Also, the
system performance in terms of computation time was
another important item of interest. Interactive work was
necessary in both test scenarios for an initial parameter
setup, the interior orientation, and the manual
measurement of control and check points. The imagery
was scanned on a PS1 scanner with 15 pm pixel size
and a standard JPEG image compression.
3.2.1 Block “Vaihingen/Enz“
This block of photo scale 1:15000 was formed by 45
images arranged in 5 forward strips (6 images) and 5
side strips (3 images) with 60 % end lap and 60 % side
lap. The block area was slightly hilly with height
differences of about 140 m covering an area of
approximately of 9.5 km by 4.5 km. Two of the forward
strips were almost identical with two other forward
strips. Thus, the block geometry was considerably strong
ape
Pr ENT
