When precise 60/60 photography is used, almost all the tie
point fall on nine different images, so the final solution is
extremely robust. When any of the three critical elements
described above is missing, the best possible approximations
are used, like digitizing the photo centers from the flight plan,
assuming zero camera angles, and assuming flat ground.
2.2 Locating control points
Also part of the initialization is an automatic control point
measurement, where this is possible.
When panels are used to mark the control points, a pattern
matching technique can be used to locate the control point
in the images. There are different approaches to solve this
problem. With the availability of very good approximate ori-
entation values for the exterior orientation, the search space
for the control points in the image is relatively small. The
above described approximate values result in a search area
which is approximately 100 by 100 pixels. Also the approxi-
mate size of the panel in the image and the panel orientation
relative to north up to 10° is known.
Figure 2: Image showing a paneled control point
Figure 3: Pattern of the control point (left) and result of the
automatic control point localization (right)
This enables us to use control point patterns and a cross
correlation technique to estimate the precise position of the
control point within an accuracy of a tenth of a pixel, which
takes less than a second per control point on an SGI Indigo2.
Figures 2 and 3 show the result of a control point localization.
In the case only natural control points, such as fence corners,
small bushes or line intersections are available, the approxi-
mate values can be used to automatically pop up all the image
sections where a certain control point falls. This speeds up
the manual control point measurement dramatically, and also
prevents misinterpretations.
3 Automatic Aerotriangulation
INPHO's MATCH-AT software is based on the same match-
ing algorithm (FORSTNER W. 86; FORSTNER W., GULCH
E. 87) as the MATCH-T automatic DTM correlation soft-
ware. The initialization phase shows the computer where to
look for homologous image features (feature points). Start-
ing at a low resolution (usually the 6-th pyramid level), and
working down in an iterative process to the highest resolution,
planned in our case to be 30 um. As the final sigma naught is
of the order of one tenth of a pixel, or 3 jm, this resolution
is quite sufficient for our needs. In fact the use of 60 um
scans may will be sufficient for creating digital orthophotos
(but not contours), but this remains to be seen.
At each tie point location a cluster of 20 - 30 points are
matched at each pyramid level, and the outliers are detected
using a robust estimation technique similar to PAT-B.
The result of MAT CH-AT is very accurate exterior orientation
elements of each photograph.
4 Automatic DTM Generation
The task here is to generate automatically a high precision
DTM in urban areas describing the topographical surface of
the ground. Topographical surface, in this sense, means the
real ground surface without objects like buildings and trees.
We use the commercial software package MATCH-T ? which
we apply to six different stereo models from six different per-
spectives showing the same area on the ground, due to the
special flight configuration.
In the next sections we first discuss the problems of the auto-
matic DTM generation in urban areas, then a new approach
for generating high precision DTM's in urban area is pre-
sented, and finally we introduce our procedure for generating
a DTM automatically for a whole block of images.
4.1 DTM generation in urban areas, a challenge for each
automatic system
The basic strategy of MATCH-T is described in KRZYSTEK
91. The principle is to use a hierarchical matching strategy
to find homologous image features (feature points). Starting
with a low resolution (top of the image pyramid), and a plane
or other external prior information as an approximation of the
ground surface, homologous feature points are used to com-
pute their 3-D coordinates and from these, a refined DTM.
This is done for each pyramid level using the DTM of the
previous step as an approximation until the highest resolution
(bottom of the image pyramid) is reached. MATCH-T uses
a regularization (ref. TERZOPOULOS 86) technique when de-
riving the DTM from the matched raw 3-D points (further
called a 3-D point cloud). This technique is designed to treat
measured points on trees or buildings as outliers and to elim-
inate them. This works well on small scale imagery, but not
as well in large scales.
There are several reasons why using automatic DTM corre-
lation directly is not appropriate for our purposes:
? Copyright: INPHO GmbH, Stuttgart
870
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
bu Mats mite ee
ns
