2. Androx ICS-400 system with 2Mbyte video
memory, an extensive library of C-callable graphics and
digital signal processing functions.
3. Two Panasonic WV-CD20 CCD cameras with a
resolution of 560 by 482 (8.8 mm by 6.6 mm image field)
pixels and changeable lenses.
4. A NEC Multisync color monitor.
3.3 General Procedure
The experimental procedure can be summarized in the five
following steps.
1. For this experiment, the control/test model consists
of ten precision machine blocks whose dimensions are
known. The size of the model is approximately 6 by 6 by 4
inches. The CCD cameras are positioned about 7 inches
apart and about 26 inches above the model. Multiple
images are taken with both cameras under various
illumination conditions.
2. Using the image displayed on the monitor for each
camera, determine the image coordinates of the control
points on the surface of the model whose object coordinates
were previously determined. For the purpose of easy and
accurate recognition, a set of well-distributed corners of
the blocks were selected. A C-language program, which
uses various digital signal processing functions of the
Androx system with operator's interactive instructions,
was developed for determining the coordinates of the
chosen points.
3. Using both cameras, images were taken of a surface
of random dots which was superimposed upon the control
model as shown in Fig. 3. The illumination should be
carefully arranged so that both cameras receive
approximately the same amount of exposure. For the
purpose of noise reduction, more than one image is taken
and averaged.
4. The images containing the random dots were
matched. The image coordinates of the dots resulting from
the matching process were then placed into the same input
data file which contains the image coordinates of the object
control points.
5. Thebundle adjustment program was executed on the
input data files using different weight constraints on
particular variables. As a result the object coordinates of
the model points, as well as those of the dots on the surface
of the specimen, were determined.
3.4 Computational Procedures
The chief computational procedures utilized during the
experiment included the self-calibrating analytical
photogrammetry bundle adjustment method and the image
matching method which was used to determine the image
coordinates of the random dots.
3.4.1. Bundle Adjustment Method. This solution was
patterned after Brown [10], [11]. Contributions to the
program were made by Orrin Long, Marquess Lewis and
Mark Nebrich. The software was modified by Weiyang
Zhou for this application.
The basis for the solution is the collinearity equations as
follows:
gu 10 Kueh ima SZ]
Fx = Xs Xpp*! | mg) Rj Kio+maolY Vid +m33(Zj Zid
Al (X;-Xic)+mo2(Y;-Yıc)+Mo3(Zj-Zie) ]
Fy 7 Ys'Ypp*f | m5; (X X.) maa(Y- Yo) &mas(Z Zi)
Where m is a function of camera orientation angles w, ¢,
and x and Xjcs Yic and Zic provide the camera's position in
object space. Xj, Y; and Zj are the coordinates of point j in
the object coordinate system. Interior orientation
parameters are represented by f, Xpp, and ypp. The image
coordinates are xg and ys.
Using a linearized version of these two equations for each
image point a least squares solution provided all of the
parameters and object coordinates after sufficient
iterations.
The project efforts are currently experimenting with the
expansion of the self-calibration techniques through the
incorporation of additional parameters which affect the
image coordinates.
3.4.2. Matching Method. Image-matching
methods fall into two groups. With the area-based
methods, such as [2] and [6], two windows of pixels, one on
each image of the pair, are judged to be a match or not
according to the similarity between the intensities of the
pixels within the two windows. The similarity is
determined by calculating statistical values, making these
methods statistical by nature. The second group of image
matching methods is based on feature, usually edge,
information of images [3] [5] [9]. The form and
distribution of the features in images are used instead of
absolute intensities of the pixels.
It is now generally agreed that edge-based methods have
advantages over the area-based methods because it is more
reasonable to match images by the variation of pixel
intensities than by absolute values of pixel intensities and it
is usually more economical in terms of computing time,
though there are some methods for improving the
efficiency of area-based methods [8].
In mechanical experiments with paper, since there is
usually not much texture on paper surfaces, it has been a
common practice to place a random pattern onto the surface
to enrich the texture. For example, dots with irregular size
and shape are used in many experiments. Inthis work, these
dots serve as targets for feature-based image matching.
In order to measure the deformation, there are two types of
matching. 1) The matching of two images taken by the
same camera before and after the deformation of the