AERIAL
Reconstruction,
rge scale aerial
ficulties for the
topographical
thed contours.
ted as supplied
ng illumination
atment as well
face is a Lam-
)ntinuous 2i-D
.. Keeping this:
reconstruction
] mathematical
tection and for
nts of postpro-
istinguish three
cannot exactly
rich only occur
7. image blun-
1g the exposure
1. We prove by
more than two
ble.
noise: Surface
tion of surface
ess can not be
s the bound to
cannot be cir-
plied by Facets
detection and
ency of the 25-
ust be approxi-
should remedy
it Facets Stereo
eliminate parts
of the surface — of course the outer surfaces of buildings,
leaved trees, parking cars, etc. become part of the recon-
structed surface.
The examples presented in this paper deal with these prob-
lems and show the surface reconstruction results of appropri-
ate areas.
3 SCENE AND DATA
We use 4 black and white aerial images for our experiments,
cf. fig. 1. They are taken by the aerial camera ZEISS RMKA
with an image format of 23 x 23cm? and a calibrated focal
length cx = 153mm. The flying altitude of 600m above
ground corresponds to an image scale mp ~ 1 : 4000. The
exposures were taken in the south of Germany at the be-
ginning of springtime, so the vegation is still leafless. The
exposure interval between the images 133/135 and 268/270
is about 20min.
Figure 1: Overlap of aerial images (mp ~ 1 : 4000) in rela-
tion to the reconstructed orthophoto
The images were scanned by the photogrammetric scanner
ZEISS PS1, with 8bits per pixel and a pixelsize of 15x 15um?.
The mathematical model of Facets Stereo Vision deals with
noisy image data — straightforward, neither geometric nor
radiometric preprocessing of the images has been applied.
The scene contains different degrees of difficulty of topo-
graphical surface types: Relatively flat agricultural areas,
steep slopes, different kinds of leaved and unleaved vegetation
and man-made objects like buildings and highway bridges.
In our experience, difficulties of surface reconstruction usually
grow within the multigrid process by growing image scale. So
we choose an image scale as large as possible to get in contact
with all of the problems of terrain noise and discontinuities
in object space. As distortions like image blunders usually
disappear at higher levels of the image pyramid, the perfor-
mance of Facets Stereo Vision concerning this point, too, can
be shown to be best at finest image resolution. Consequently,
in this paper we only present results gained with the original
pixel values and resolution from scanning.
4 PARAMETER SETTINGS AND SOME SPECIAL
HINTS
We started the multigrid process at the 9th level of the image
pyramid. The corresponding resolutions are given in tab. 1.
759
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Start values for the heights were obtained by simply bilinear
interpolating the coordinates of four outlying tiepoints from
the bundle block adjustment.
level: | pixel /facet: size:
image pixel 4 x 4mm? in image space
orthoimage facet 32 x 32m? in object space
9 ~ 2 X 2 image pixels
height facet 128 x 128m? in object space
~ 8 x 8 image pixels
image pixel 15 x 15um^ in image space
orthoimage facet 12.5 x 12.5cm? in object space
1 ~ 2 X 2 image pixels
height facet 50.0 x 50.0cm? in object space
~ 8 x 8 image pixels
Table 1: Facet parameters and image pixel sizes for highest
and lowest multigrid level
The test area covered ~ 600 x 600m? in object space, cor-
responding to ~ 1.4 - 10° estimated height parameters and
~ 22.4-10° estimated orthoimage grey value parameters. Be-
cause of the high image noise of go ~ 46 — 8 grey values we
choose non adaptive curvature minimization as regularization
procedure with a weight factor of À = 1 - 10°. The relativly
large size of the heigt facets in relation to the pixel size causes
some implicit regularization, too.
In principle, the long time delay (cf. section 3) between the
two flight strips can be taken into account by simultaneously
estimating one separate set of orthoimage parameters for the
images of each strip. The basic idea of this proceeding is very
similar to the treatment of color images, as explained by [1].
By that way, the different surface texture caused by different
shadow locations in the images of the different strips can be
taken into account precisely. In this paper we treat all images
as isochronous exposures. So, small errors in those regions,
were shadow gives texture, have to be expected.
The contours plotted into the orthoimages on the following
pages of this paper exactly reproduce the results obtained by
Facets Stereo Vision by bilinearly interpolating within each
height facet for every orthopixel position. Please note, that
our goal for this paper is to document the original reconstruc-
tion results of Facets Stereo Vision, but not the results of any
additional fine contour smoothing algorithm!
We believe that it might be useful not only to offer height and
orthoimage data, but also the accompanying quality criteria
to a further semantic analysis. This in mind, the decision
whether to build a 'good looking' smooth result or not should
be dependent on the aim of the following data procession
steps.
5 ACCURACY CHECKS
We obtained the parameters of the outer orientation of im-
ages by bundle block adjustment based on image coordinates
measured in the analogue images. The bundle block ad-
justment reached an accuracy of +3cm for the height com-
ponent of the tiepoints in object space. To ensure that
the set of our transformation parameters is really correct
for the digital images, we compared a set of 10 signalised
control points with the corresponding interpolated heights
of Facets Stereo Visions surface reconstruction: The abso-
lute height differences all were less than |AZmaz| < 10cm,