t level with the
le procedure is
1e, called point
level until a so-
t is called point
ing levels. The
el is to arrive at
vailable amount
'd matching is
. For each level,
y in each image
then matched
ometric criteria,
ite points. These
ed into a robust
rmines both the
ige pair and the
onjugate points.
luring the robust
sistency check.
ree-dimensional
led to the next
feature-based
stops at the
point matching
the image pair
f the conjugate
el.
acking, a fine
f the conjugate
'el is conducted
cermann, 1983)
Around a given
eference and a
imeters and two
1e two windows
pair, the cross
vo surrounding
nt is larger than
ful. The interest
indow to find a
' pyramid level.
rch window via
'sponding point
the next lower
At the end of
ly tracked to the
robust bundle
final relative
and the three-
oints. The point
tracking is of great advantage to speed up the whole
procedure without suffering any loss in accuracy and
reliability of the results. It ensures that the search for
conjugate points is done only in areas in which well-
defined features can be expected.
3. IMPLEMENTATION
In addition to speeding up the algorithm while
maintaining the reliability of the result, special attention
was paid during the implementation of ARO in PHODIS
ST to
- restrict user input parameters to a very limit, and
- keep the number of conjugate points in the image
pair to a reasonable number.
For point matching a number of control parameters such
as window sizes and threshold values exists. Their optimal
setting changes with different kinds of image texture,
scale and terrain types. In order to avoid the parameter
setting by users, a number of different sets of control
parameters is used for point matching on every pyramid
level. This has the advantage that the control parameters
can be adapted to the image material. However, the
computation time increases if runs with multiple
parameter sets are performed one after the other.
Therefore, all the different parameter sets are used in a
one-pass operation in the current implementation. This
leads to a decrease of the computing time while
maintaining the reliability of the results.
Approximate overlap values of an image pair were
optional parameters in the early algorithm. The current
implementation includes a function that determines the
overlaps automatically, so that a user input is no more
necessary. First, the feature-based matching is performed
assuming the overlaps to be 80% end overlap and 100%
side overlap, however with a large tolerance. Using the
matched point pairs, a robust least squares adjustment of
the x- and y-parallaxes is then conducted. In this way,
outliers in the matches are also eliminated to some extent.
The adjusted parallaxes approximately represent the base
components in image space and overlap values can
directly be derived from these values.
Usually, the number of conjugate points determined by
the automatic procedure from an image pair can be
unnecessarily large for the computation of the relative
orientation parameters. Therefore, the current
implementation tracks conjugate points only selectively.
A grid is used in the overlapping area to control the point
selection. This also speeds up ARO considerably, since
point tracking needs a lot of I/O operations and is thus
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
very time-consuming.
In summary, the current implementation of ARO in
PHODIS ST requires no user input parameters except the
order of images, because the camera data and the pixel-
image coordinate relationship belong to the standard
PHODIS image. The preparation steps of ARO are:
- define the left and the right image,
- check whether the image information is complete,
eg. camera, interior orientation and pyramid,
- if not complete, call corresponding tools to accomplish
it,
- if complete, start the procedure.
4. OPERATIONAL TESTS
In order to assess ARO for the photogrammetric practice,
the algorithm was tested with 53 image pairs. The image
pairs differ in pixel size (12.5-30 um), ground cover
(rural, forested, urban, glacial, desert), and terrain type
(flat, rolling, mountainous). Image scales range from
1:3,200 to 1:54,000.
Table 1 contains a classification of 47 image pairs into 9
groups according to image scale and ground cover/terrain
type. Moreover, 6 special cases have been investigated in
order to find out the limits of the developed approach.
These are characterized in Table 2.
In the following the test procedure is described. First the
analogue images were scanned, mostly with a pixel size
of either 15um or 30um and 8 bits per pixel. In the next
step image pyramids were generated. Then, the interior
orientation of the images was determined using either the
automatic module AIO (Automatic Interior Orientation) of
PHODIS ST, or interactive measurements. Then, ARO
was started. No parameters whatsoever had to be provided
for ARO. For verification purposes, epipolar images were
computed using the orientation parameters from ARO.
Finally, the epipolar images were checked for remaining
y-parallaxes by stereoscopic viewing in the PHODIS ST
environment.
In the sequel the results of the successful ARO runs are
discussed with regard to remaining y-parallaxes in the
stereomodel, accuracy, reliability and computing time. The
main focus is to analyze the accuracy, represented by the
variance factor O, a posteriori, as a function of image
scale, ground cover, terrain type, pixel size, overlap,
number of conjugate pairs and image quality. While ARO
was successful for all stereo pairs of Table 1, the
procedure failed in three of the special cases of Table 2.
These are discussed towards the end of the chapter.
