failure.
gauges.
surface
surface.
ucture’s
mber of
lence is
ing load
f digital
jatching
through
mmetric
een the
se-range
physical
earth's
rformed
ed using
image
immetric
d work-
isticated
a small
| relative
ion with
raphs of
r terrain
nce from
ges and
enerated.
partially
y density
xXtographs
r second.
carry out
» on the
gue and
ation of
n marked
a highly
International Archives of the Photogrammetry, Remote Sensin
automated procedure. Examples of it being carried out
are abundant in the literature, examples are given by
Whiteman ef. aL (2002) & Li and King (2002).
However, this paper is concerned not so much with the
finding of targets as with the measurement of surfaces,
defined by numerous points found through the automated
matching. Although this can often be carried out using
the same work-stations intended primarily for aerial
photographic use, the peculiarities of non-topographic
measurement often mean that it is carried out using
cheaper proprietary software loaded onto personal
computers or work-stations or software developed in-
house such purposes.
One of the most important aspects of carrying out
automated measurement on either aerial or non-aerial
imagery is the "digital image matching", or “image
correlation" procedure. Over the last few decades, a
number of approaches have been developed but generally
these techniques can be classified into two main groups,
feature-based and area-based matching. The former
group is fast and reliable and capable of finding matches
with poor initial values, while the latter approaches have
the advantage of high precision. It is the latter group
which is of interest here.
Area-based stereo image matching techniques make use
of two small areas (windows) surrounding the point for
which matching is needed within each image. A
correlation technique, generally based on least squares
estimation, selects the point of best match. Methods in
area-based matching have developed since early
significant, seminal work by Foerstner (1982) and Grün
(1985). Further significant extensions of area-based
matching were proposed by Grün & Baltsavias (1987);
Rosenholm (1987a & 1987b) proposed a method of
multi-point area-based matching technique in evaluating
three-dimensional models, and area-based method was
further extended by Baltsavias (1991) through the use of
images from several viewpoints, i.c. with multiple
images. Further developments of the area-based method
was proposed by Wrobel (1991) and Heipke (1992) in
which the matching integrates image matching and
object surface reconstruction. Recent developments,
among others, was proposed by Di Stefano et.al. (2002)
whereby a fast area-based stereo matching algorithm was
proposed.
3. AREA-BASED IMAGE MATCHING MODEL
USING A SURFACE MODEL
Mathematical details of conventional area-based
matching (ABM) are provided in many publications and
are not given here. It may be sufficient to point out that
ABM is based on least squares solution of a number of
similar equations, one being written for each pixel in the
selected “window” surrounding the point to be matched
on one image. The equation for any pixel incorporates
the image co-ordinates in one window and the image co-
ordinates of corresponding pints in the other window, via
a transformation equation. Conventionally, an affine
transformation between the windows is adopted,
(Foerstner, 1982). The unknowns to be determined in the
least squares solution are the parameters defining this
transformation - including parameters which define the
g and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
659
relative positions of the windows. In effect, these
parameters represent what are conventionally called x
and y parallaxes. The equations often also incorporate
radiometric values but these are not significant to this
brief explanation. No information of the object is taken
into account in the matching process. Matching is solely
based on the intensity values of the pixels and the
assumed affine transformation.
An attempt has been made to improve the accuracy of the
traditional area-based technique as offered by Grün
(1985). The revision involves replacing the conventional
model which is used to transform one window shape to
the other to improve the mathematical description of the
relationship between the windows in the least squares fit.
The new transformation incorporates a simple model of
the surface being measured, and replaces the assumption
that the windows differ according to an affine
transformation. lt serves as a compromise between the
traditional and the far more complex global area-based
matching method. As with the conventional solution, the
method proposed here solves, through an iterative least
squares solution, for the corrections to image co-
ordinates (x,y) of the search window. But, in addition,
two ‘new’ unknowns, the gradients in X and Y directions
on the surface at the point on the surface which
corresponds to the centre of the search window and their
second derivatives are introduced.
Since the transformation used is more rigorous than the
affine, it is hypothesised that the improved functional
model will allow the use of larger windows for matching
and hence improve accuracy. It is also found that the use
of a better functional model will converge more quickly
to give a solution.
The mathematics of the refinement can be obtained from
another more detailed discussion, (Mustaffar & Mitchell,
2001). It may be sufficient to point out that the model
described above is extended by taking into consideration
the shape of the object. The equations used to define the
transformation now includes some information about the
object’s surface which needs an additional significant
complication of using the known relative positions of the
cameras.
4. EXPERIMENTS
Tests of the algorithms were carried out in the study of
steel deformations under static loading, see Figure.
Images were taken using a pair of Kodak digital still
cameras which comprise of a DC290 and DX4900. The
DC290 is fitted with a 6.0 mm lens and has a resolution
of 3.1 megapixles. The lens and resolution for the
DX4900 are 7.3 mm and 4.0 megapixels respectively.
These cameras are mounted with a base distance of
approximately 750mm and the relative orientations of
these cameras has been determined. A projector was
placed in between the cameras for the purpose of
projecting patterns onto the object. The pattern used in
this investigation was a diamond shaped mesh.