specimen. 2) The matching of two images taken by the two
cameras respectively at the same time. With type 1)
matching, the difference of the shapes of the same dot in
two images was mainly caused by the deformation. With
type 2) matching, the difference is mainly introduced by the
angle of projection. However, viewing a dot in the image as
a density distribution function, the transformation of one
function into the other is mainly one of the translation of
both X and Y axes. Therefore, the characteristic parameters
used to describe a dot must be, first of all, translation
invariants. In this experiment, central moments of the first
four orders of each dot were calculated as the dot's
signature.
Hoo-7moo. M10-Ho1-0
Hao = M90 - Hoox?
611 = M11 - HooxXy
Ho2 = Mo2 - Hooÿ*
30 = Mgo - 3m90X + 2p00x°
os = Mo3 - 3mo2ÿ + 2Ho0ÿ*,
Jio] 7 mo] - mggy - 2m11x * 2uo0x2y
119 7 m]9 - mogy - 2m11y * 2uo0xy?
400 12°
here, Hpq7 | J(x-X)Pty-y)8f(x.y)dxdy
-00
co tee
mpg = j [xPy f (x. y) dxdy
-00
and x-mjo/moo, y = Mo1/M00
In addition to these, size is an important part of the signature
of a dot. However, during the experiment, the size, as well
as other factors of the signature, changes from image to
image, due to both the projection angles and the
deformation of the specimen. Fortunately, a valid
assumption of this experiment is that the "distortion" of the
shape of a dot in two images is always moderate within
small intervals of time between image taking during
deformation, and with almost identical projection angles,
i.e., with the two cameras kept as vertical to the specimen as
possible. Therefore, though they are not exactly invariants
in all the images, this set of characteristic values, including
central moments and size, comprise a reasonably good
signature of a dot.
Because of the digital nature of the images, it is easy to see
that dots with a larger size have richer signature
information. For example, in the extreme case, dots
consisting of a single pixel are all the same in shape. For
this reason, as the first iteration of image matching, the
biggest dots in both images are selected and matched first
by their signatures. However, as discussed above, the
signature factors are not exactly invariants. As a
consequence, the matching obtained only by signature is
not absolutely dependable. An angle test is carried out for
the matched dots in the first iteration by checking the
difference between the angles formed by any three dots in
the first image and the angles formed by the three
correspondingly matched dots in the second image. If the
difference is not smaller than a predefined threshold, the
matching of one pair of the matched dots will be judged as
unacceptable and deleted from the list of matched dots. The
angle-test for the first iteration must be done very strictly
since the matched dots are going to serve as "seeds" in the
following iterations of matching.
After the first matching of the largest dots in the image, a
window of adjustable size is opened for each of the two
matched dots in each image respectively. Other unmatched
dots which fall into the window in two images are matched.
Asin the first matching, at the end of all the dots in a pair of
windows, the "angle-test" is carried out to delete the
matches with significant difference in the angles. The
threshold to determine if the angle difference is tolerable is
dynamic. The more similar a pair of dots are in terms of
their signature factors, the looser the threshold will be. In
other words, the pair of dots that are very different in their
signature have to pass very strict tests on the angle they
form in order to be a match.
The dots which are matched in the second matching in turn
serve as "seeds" in the third iteration of matching, and so
on. The matched dots are no longer considered in the
following iterations. The entire matching process comes to
an end when all the possible pairs of dots are matched.
After the entire matching process, only the heights at the
dots, or more precisely, the heights at the center of the dots,
will be solved accurately with intersection. The heights of
other points, if of interest, will be determined through
interpolation.
After matching the dots in the image, the image coordinates
are put into the input data file for a least squares adjustment.
4. RESULTS TO DATE
Sample results are listed in Table 1. These results are for a
two camera station solution which utilizes three types of
object points, namely, object control points, object check
points, and object random dots which will define the final
surface deformation of the specimen.
The determination of the interior and exterior orientation
parameters for each camera were based on seven (7) known
points which were located on the surface of the model. For
these points the average absolute residuals in inches
between known and final adjusted X, Y and Z coordinates
were 0.0024, 0.0021, and 0.0006. An additional eight (8)
check points were included in the solution. The weights of
these points were zero and thus the solution was free to seek
the best fit coordinates for these points. The average
absolute residuals in inches in X, Y and Z for these points
were 0.0295, 0.0301, and 0.0517.
There were no checks on the computed coordinates for the
132random dots; however, the average standard deviations
in X, Y and Zin inches were 0.0062, 0.0062, and 0.0269 for
the listed results in Table 1. In general theresults in X and Y
are better than in the Z direction.