NA
7%
?
£2
M
M
x
25
©
9,
rl LS
7,
area.
orientation
the area in
med to be
1/technical
easy target
l. For easy
' structures
je used. If
successful,
ke manhole
om T2/T3
nd parking
| map. It is
Itersections
at it is pos-
t template.
anual edit-
s available,
interior ori-
es by using
flight plan,
each image
/ chosen to
system. In
> template-
1e 100, and
An example is shown in figure 2.
Figure 2: Target template (road intersection from the lower
left of figure 1).
Splines are treated as straight lines between the registered
points for reasons of computational speed. This introduces
systematic errors for non-symmetric road intersections. The
solution to this problem would be to compute intermediate
points.
3.3 Image pyramids
A coarse-to-fine strategy is implemented by using image pyra-
mids. The original, high-resolution image is usually referred
to as level 0. In the next pyramid level, level 1, the pixel
size is twice the pixel size in level 0. In this study, the pixel
values in level 1 are simply computed as the mean values of
the corresponding four values in level 0. Image pyramid levels
2 to 5 are built in a similar manner.
With the test data available (please refer to section 4!), there
seems to be no reason to proceed any further than level 5: In
level 6 even large structures like roads are fading.
3.4 Matching algorithm
The matching algorithm is designed to be running completely
automatic, providing solutions to both the detection and the
pointing problems by means of high redundancy in the data
and a coarse-to-fine strategy. For the bundle adjustments
and visualization of results, however, we have used the DEM
program package MATCH-T from INPHO GmbH. The algo-
rithm:
e For image pyramid level 5 down to level 0
1. Project closed polygons for (road) intersections
derived from a digital map into the images, us-
ing the orientation parameters available. For the
first iteration, read approximate orientation pa-
rameters from a prepared file.
2. Sample the target templates, using the nominal
pixel size for the actual pyramid level. For each
target template, one of the polygon points is cho-
sen as the template origo.
3. Compute the correlation coefficients in an ap-
propriate search area. For the first iteration, the
search area must be large, 4x4 cm?, in the image
scale.
4. Determine subpixel image co-ordinates for all tar-
get templates by computing a least squares fit to
a bivariable second-degree polynomial.
5. Compute improved image orientation parameters
by means of a robust bundle adjustment.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
e Proceed to smaller ground control objects for final
measurements and orientation parameter computation
(this has not yet been investigated!).
In the bundle adjustments using measurements of large con-
trol objects, a priori standard deviations for horizontal and
vertical control are set to twice the standard deviations for
photogrammetrically well-defined points, according to the
specifications. A priori standard deviations for image points
are computed as the RMS of all subpixel estimations. This
may not be entirely realistic, but enables automatic iterations
down through the image pyramid.
4 TEST DATA
4.1 Image data
The image data used for an empirical test have been pur-
chased for educational purposes, and are therefore easy at
hand at the laboratory. Two panchromatic photographs with
a 6096 forward overlap at scale approximately 1:5,000, taken
with a Zeiss camera of type RMK-A15/23 (czz153 mm) have
been digitized on a Zeiss/Intergraph PS1 scanner, with 8 bit
radiometric resolution and 15 um pixel size.
In the model area, 50% or less is a suburban residential area
and the rest is covered by agricultural fields. A small fraction
of the urban area has been built recently at the time the
photos were taken.
4.2 Map data
The map data available is a digital technical T3 map from the
local municipality. The map has been revised since the pho-
tographs were taken. From this map 30 (road) intersections
consisting of 566 registered points have been extracted. Also,
47 arbitrarily selected from 240 available manhole covers in
the same urban test area have been extracted for reference.
In figure 3 (left), the model area is shown together with the
manhole covers which are used as reference points. In figure
3 (right), the positions of the extracted intersections are in-
dicated. As can be seen, the test area covers less than 50%
of the model area.
Figure 3: The model area with the manhole covers used for
manual reference measurements (left), and with the positions
of the extracted intersections indicated (right).
