HIGH RESOLUTION SURFACE RECONSTRUCTION OF A LANDSCAPE FROM LARGE SCALE AERIAL
IMAGERY BY FACETS STEREO VISION - AN EXTENDED TEST
M. Schlüter and B. P. Wrobel
Institute of Photogrammetry and Cartography
University of Technology Darmstadt
Petersenstr. 13, D-64287 Darmstadt, Germany
wrobel@ac3.phgr.verm.th-darmstadt.de
ms@ac3.phgr.verm.th-darmstadt.de
International Society of Photogrammetry and Remote Sensing, Commission Ill, Working Group 2
XVIIIth ISPRS Congress, Vienna, Austria, July 1996
KEY WORDS: Photogrammetry, Vision Sciences, Matching, Orthoimage, DEM/DTM, Geometric Surface Reconstruction,
Stereoscopic Aerial Imagery, Facets Stereo Vision
ABSTRACT
We present results of high resolution surface reconstruction in object space by Facets Stereo Vision, using large scale aerial
images. After some references to the basics of Facets Stereo Vision, we try to classify possible sources of difficulties for the
surface reconstruction. Consequently, we choose appropriate examplary areas in object space with different topographical
character. For these areas, the results of Facets Stereo Vision are given in form of orthoimages and unsmoothed contours.
Further on, corresponding quality statements like residual images and images of standard deviations are presented as supplied
by Facets Stereo Vision.
Topics like the occurrence of image blunders, moving cars in object space during the exposure interval, changing illumination
during the exposure interval, terrain noise and discontinuities in object space are covered. We discuss their treatment as well
as the reached accuracy.
The results seem to be promising for further scene understanding tasks.
1 ABOUT FACETS STEREO VISION
Facets Stereo Vision is a method for surface reconstruction
in object space, cf. [6]. DTM-heights, orthoimage grey val-
ues and radiometric parameters are directly introduced as un-
known parameters, and are estimated simultaneously by least
squares adjustment from two or more images containing one
or more spectral bands.
For the examples in this paper, the relationship between
image space and object space is described by the perspec-
tive camera model. Both, DTM-heights and orthoimage
grey values, are represented as 21-D functions over regular
grids (called facets), using bilinear interpolation between grid
points. Regularization, as discussed in cf. [7], is required to
overcome ill-posedness of the image inversion and to bridge
areas, in which the gradients of the image grey value signal
are low in relation to image noise. As least squares adjust-
ment requires Taylor-linearization, approximate start values
of the parameters are needed. Therefore, the whole recon-
struction process is formulated as a multigrid procedure: Step
by step, a finer resolved DTM is calculated for each level of
the image pyramid of the input data. For details see [5]. A
consistency check with respect to occlusions is useful to avoid
the matching of pixels to hidden parts of the surface.
Our aim is not only to reach a surface reconstruction fitting
tightly to the real world surface, but also to gain correct
quality statements for our results as well.
2 WHICH RESULTS CAN BE EXPECTED?
Contrary to an operator with an analytical plotter, who mea-
sures single point positions and maybe using explicit knowl-
edge for point selection, Facets Stereo Vision works area
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
based, assuming e.g. that the real world's surface is a Lam-
bert reflector and can be approximated by a continuous 2i-D
function over regularly spaced grid points, etc. . Keeping this:
in mind, it is clear that problems of the surface reconstruction
have to expected at those areas, where the used mathematical
model does not fit properly to the real world.
For a brief discussion of error avoidance or detection and for
considerations concerning further developments of postpro-
cessing algorithms, it seems to be helpful to distinguish three
classes of possible error sources, even if they cannot exactly
be kept distinct from each other:
First, let us have a look at disturbances, which only occur
locally in a minority of the used images: E.g. image blun-
ders, objects which move in object space during the exposure
interval or local light spots by total reflection. We prove by
two examples (cf. fig. 2, 3) that simply using more than two
images simultaneously can avoid a lot of trouble.
A second group of errors is caused by terrain noise: Surface
elements, which would require a finer resolution of surface
facets than used in the reconstruction process can not be
approximated properly. Image resolution sets the bound to
the resolution in object space. So this error cannot be cir-
cumvented easily. But the quality criteria supplied by Facets
Stereo Vision offer powerful possibilities for detection and
further semantic analysis.
The third class of errors is caused by a deficiency of the 25-
D surface model: Vertical surface areas can just be approxi-
mated by steep facets. A 3-D surface model should remedy
this situation, cf. [4].
Last but not least it should be mentioned, that Facets Stereo
Vision does not use any explicit knowledge to eliminate parts
758
leav
stru:
The
lems
ate
cf. f
with
leng
grou
expc
ginn
expc
is ab
Figu
tion
The
ZEIS
The
nois)
radic
The
grap
stee[
and
In où
grow
we c
with
in ol
disar
mans
be st
in th
pixel
We s
pyrai
