ons.
Jital
tion
Yoto
tem
AR.
ical
ling
tion
atly.
ed.
The
s of
red,
ited
"ital
, Or
78;
han
f P.
PW
| as
etry
ical
ons
atic
rial
age
atic
age
and
ars,
ital
and
zate
ing
ing,
t is
1ost
age
Image matching is an important basic step in the three
dimensional data reconstruction problem. Based on the theories
and methods of pattern recognition, this paper discusses the
weaknesses of traditional image matching algorithms. That is, the
results from traditional image matching algorithms are
inharmonious and unreliable because they do not take spatial
relationships into account when they do not use neighbourhood
matching to adjust the global matching results. In this paper two
algorithms, probabilistic relaxation and the Hopfield model in
neural network techniques (which are used frequently in pattern
recognition) are discussed in detail with respect to their
application in image matching. Relaxation processing is a useful
technique for using contextual information to reduce local
ambiguity and achieve global consistency in global image
matching. It is basically a parallel execution algorithm. On the
other hand, the neural network is a computational model with
massively parallel execution capability. Being applied in
gray-level based image matching, they give consideration to
spatial relationships and as such, global consistency has been
improved greatly. There exist certain common properties between
relaxation processing and the neural technique. A mapping
method that makes the Hopfield net perform relaxation processing
is proposed in this paper. One advantage is that relaxation
processing can be performed in real time since the Hopfield net
can be implemented by conventional analog circuits. Based on the
new concept of zero matching, not only the correct matching
result can be contained, but also break-lines and occlusions can be
processed. Because of these aspects, the stability of image
matching has been greatly enhanced.
Not only is VirtuoZo one of the most exciting and innovative
DPW available today, it is also a superb visualisation tool.
VirtuoZo has been used on many projects, such as mapping,
management of land, scaling Australia's Ayers Rock, corridor
measurement for trains, resource management, measurement of
dinosaur tracks, highway and railway design, etc. VirtuoZo can
process not only aerial images, but also SPOT images and close
range photography.
2. FUNCTIONS AND THEORETICAL FOUNDATION OF
VIRTUOZO
VirtuoZo, which was written in C, X-windows/OSF Motif and
OpenGL (for the stereo graphics) is a Unix based application.
VirtuoZo accepts scanned stereo photography, stereo SPOT
satellite imagery or scanned close-range stereo photography as
input and very simply and quickly produces DTM, ortho-rectified
imagery (orthoimage), and contour maps as it's primary output.
Secondary output consists of dynamic three dimensional
visualisation techniques in stereo of the modelled object, and
digitised vector data.
2.0 Input
VirtuoZo processes standard 24-bit colour and 8-bit black-white
digital images. TIFF, IRIS RGB, SUN Raster, BMP and JPEG
graphics files can be converted to the necessary format from
within VirtuoZo. Scan resolutions between 7 and 200 microns
have been tested, yielding excellent results. The parametric input
through the VirtuoZo tools menu includes control data, camera
calibration data, block and model data.
425
2.1 Interior Orientation
The interior orientation in VirtuoZo is predominantly an
automatic procedure of recognition and location of fiducial marks
on digital images from the metric camera. For each kind of
fiducial mark, the system has to learn only on the first processed
image under the guidance of the operator by simply pointing and
clicking two fiducial marks using the mouse. The features inherent
in the fiducial mark will be analysed and stored, and used in the
pattern matching of subsequent fiducial marks. Semi-automatic
and manual measurements are also available. In the semi-
automatic mode, an approximate position of the fiducial mark is
pointed out by the mouse and the accurate position is located by
the system. In the manual mode, the accurate position of the
fiducial mark is pointed out by the mouse and then optimised very
simply by incremental pixel shifts. The results are displayed in
real time.
2.2 Relative Orientation
In VirtuoZo's automatic relative orientation, distinct points are
selected, and their conjugate points are matched automatically.
The feature extraction operator, the feature location operator
(Zhang J., 1992), area-based matching, local multiple point
matching and least square matching are applied based on the
dynamic image pyramid. A forced condition is used in standard
positions to ensure necessary points are acquired. The interactive
actions in both semi-automatic and manual modes enable the user
to add, optimise and delete user-defined related points, if required.
Points selected by the operator can be made on either the left or
right image. Conjugate points will be automatically matched by
an area-based matching method in the other image automatically,
or may be selected by the operator manually. The entire overlap
of the image pair is displayed on the computer screen for the
overview, and a high resolution image window is popped for
optimal selection after the approximate position is pointed out.
The results are calculated in real time and displayed to allow the
optimisation of individually related points.
2.3 Absolute Orientation
The absolute orientation is effected by the user interactively
nominating control points by way of unique identification, point
and click methods, similar to the interactive actions of the relative
orientation. Optimisation of the absolute orientation is performed
using a point and click technique to introduce incremental shifts
of one fifth of a pixel. The results are calculated and displayed in
real time.
2.4 Automatic Aerotriangulation
The tie points for aerotriangulation are automatically selected
using the feature extraction operator after only one pair of
conjugate points is measured interactively in the side lap of two
images in each of two neighbouring strips. Four images can be
displayed on the screen for the interaction, two are from one of the
strips and the others are from the neighbouring strip. Both the
images and the strips can be selected. The tie points, selected in
automatic mode, are located in three columns on each image, the
middle, the left and the right, if the conjugate points exist, and
there are at least five points in each column (excepting in low
contrast regions, such as water). Tie points are automatically
transferred by using local multiple point matching based on
dynamic image pyramids, and a coplanar restraint condition is
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
