ame known
jon. Precise
introduced
the number
fines some
jon has not
utomatically
have to be
al case of
J tools can
ration is an
cess steps
positions in
'ange of the
ing a pull-in
1 matching
the kernel
ould not be
more than 1
put data in
NS data are
1 0.2 - 0.5
y derive the
1 cm in the
, a DEM is
itical relief
become a
tractive and
DEMs are available in highly industrialized countries, we
also have to envisage a method which operates mainly
with the images and uses the additional data mentioned
above, wherever possible. The key idea is to derive a
crude block DEM at a coarse pixel resolution. If the
projection centers are known with an accuracy of 1 cm in
the images, we simply apply a DEM generation after the
automatic block adjustment based on the kernel system.
This is accomplished hierarchically in several coarse
pixel resolutions, typically in the image pyramid levels of
1mm and 0.5 mm pixel size. Eventually, the block
adjustment and the subsequent DEM generation is
applied iteratively at the same pixel resolution, until a
convergence of the Gruber positions is reached. This
method works well in any image scale provided that GPS
observations for the projection centers are given. Even in
small image scales a flight index map, which provides
the projection centers usually with an accuracy of 100 -
200 m, may be used. The only critical case remains at
large image scales with height undulations of up to 20%
and more of the flying height. In such cases we envisage
also a method doing an automatic relative orientation at
the pyramid levels of 1mm and 0.5 mm pixel size. This
method is very similar to Schenk (1995) and is applied to
all image pairs with some constraints for small overlap.
It aims at an accuracy (=sigma naught) in the final block
adjustment of 1 pixel, which means about 0.5 mm. It
remains to be seen how flexible the integrated DEM
generation approach really works if the image overlap is
known rather inaccurately for some reasons. It should be
recalled, once more, that the GPS technique has today
already reached the status of a standard, and hence only
a minority of triangulation projects will not have GPS
observations for the exposure centers in future.
57 digital images at 5
480 um pixel size
index map - GPS/INS data
azimuth - initial DEM
flying height inr
determination of tie point areas at Gruber point positions
* direct determination
from initial block data
and GPS/INS/DEM
* accuracy: « 1 cm
* DEM generation
* block adjustment
* automatic relative orientation
* accuracy: c» 7 1 pixel = 0.5 mm
- orientation par.
- coarse block DEM
- tie point areas given
in object space
Figure 2: Initialization of MATCH-AT
The main result of the initialization are the tie point areas
with an accuracy of at least 1 cm in the image.
Furthermore, a crude block DEM and image orientations
are calculated (Figure 2). The result of the initialization
can optionally be checked and edited in critical cases.
However, this interactive step is not mandatory at all
and should be reserved if the initialization reports
unacceptable results.
2.4 Kernel system
Once the initialization has completed, the kernel system
is subsequently invoked using the initialized data.
Additionally, the GPS observations for the exposure
centers are used with appropriate standard deviations.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
The image pyramid is optionally derived only at the tie
point areas in advance. This sparse image pyramid
structure reduces the amount of image data to some
extent, although an image compression like JPEG is
certainly more effective.
The workflow and matching scheme of the kernel system
shown in figures 3 and 5 is applied in all tie point areas.
Firstly, a preliminary matching in all image patch
combinations yields a list of image point pairs. Those
points represent only two-fold points and are tied up to
many-fold points by means of a heuristic search
procedure. Since especially in the coarser image pyramid
levels the percentage of erroneous preliminarily matched
points is considerably high, the result of the heuristic
search has to be considered error-prone. The subsequent
bundle solution, however, eliminates iteratively the
mismatches by using robust statistics and estimates
simultaneously orientation parameters for all images.
This is a remarkable feature of the kernel system
compared to other approaches. It uses rigorously the ray
intersection as a geometrical constraint in the matching
of multiple points. Thus, the approach is not limited by a
geometrical assumption about the local matching area
(e.g. plane), but provides also matched points on corners
of houses, for instance.
After the block adjustment the DEM is updated either at
the tie point areas or for the entire block. The local DEMs
at the Gruber positions are useful to resample the image
patches, which are to be matched, with respect to the
terrain surface. The area of the local DEM is slightly
larger than the matched image patch in order to
overcome edge effects. The knowledge of the terrain
surface is advantageous especially in hilly or
mountainous terrain and permits the use of larger image
patches, as long as the DEM fits accurately enough the
terrain. We use the same surface reconstruction
technique as in MATCH-T which operates with a robust
finite element technique (Krzystek, Wild, 1992). The grid
width corresponds to approximately 30 pixels. The DEM
generation process takes full advantage of the multiple
image overlap. If compared to a conventional automatic
DEM generation with two images the multiple image
DEM approach creates more terrain points per grid mesh
= square of 4 DEM posts). Additionally, the terrain
points are intersected by more than two rays, thus,
increasing accuracy and reliability of the DEM.
After the DEM generation, which is practically an option
of the system, the tie point areas are updated, using the
terrain surface, if it was determined. The described
scheme of the kernel system is applied straightforward
through the entire image pyramid and results in adjusted
object coordinates and image orientations parameters.
2.5 Summary
The approach integrates the point selection, the point
measurement, the point transfer and the block
adjustment in one single process. Instead of single tie
points clusters of points are created. Those clusters are
tracked through the entire image pyramid (Figure 4).
Thus, we do not track hierarchically single features
through the image pyramid which might get lost. Instead,
407
————————————
