sed on the
erse matrix
M.S. of the
measured
| the same
OINTS
e images,
een image
dentified. If
at least in
-parameter
ts in each
all targets
flight plan
ations may
pipolar line
ay be used
ecause it is
| values for
2 search for
vipolar line,
ularly when
troreflective
area should
ds, epipolar
least three
able a priori
should be
amount of
iere human
tasks easy
In practice,
perator will
en will label
numbers or
. 2). This is
s and many
up a semi-
job will be
ver it stops
. The other
matically.
sed
arget image
the control
points; the operator looks for pairs of images containing
at least the same 6 targets, 3 of which being control
points, and begins by pairing all these image points. The
(asymmetric) relative orientation for each pair is
computed, starting from approximate values of the
rotations and base components. In our case, since the
images were taken in pairs with two cameras mounted
on a rod, the orientation of the images of a pair is easy,
more difficult otherwise. Rather than using the Cardan
sequence for the rotations, we express their approximate
values in terms of a Euler sequence, which are easier to
provide, and solve from the rotation matrix with respect
to the angles w,ÿ,x. The models are thereafter absolutely
oriented: since the Anblock hypotheses are not satisfied,
we compute the transformation parameters by
expressing the rotation matrix by the Hamilton's
quaternions (Sanso, 1973) which do not require initial
values. Finally, a bundle adjustment provides the exterior
orientation of the images. Since the distribution of the
targets in the images may not be ideal, the adjustment
will be repeated at a later stage, when more image
points are made available. If the initial configuration is
too poor, nevertheless, the procedure may not continue
correctly, due to a lack of targets in the area.
We can now take advantage of the known epipolar
geometry selecting from the pairs sets of three images,
to find automatically more corresponding targets
(Ayache, 1991; Maas, 1992).
Let A be the reference image, B and C the other two
images. For each target on A, the corresponding epipolar
lines eg and ec are drawn on B and C respectively;
targets Pg; and Pc, falling in a band along the two lines
are considered (see Fig. 3).
Starting from a point Pg; in B its epipolar is drawn on C
and its intersection with ec is computed. If a candidate
Pc, is found in a window defined by the two bands
overlap around this point, then a candidate set is found;
more candidates Pc, may anyway exist in the same
window. Once all possible sets are found, ambiguities
must be solved through a consistency check of the
intersecting rays. This is done by computing, for each
set, the three distances between the rays and taking their
mean value. The set enjoying the minimum distance d,
is taken. If din is smaller than a threshold, the ground
coordinates of the point are computed. Once all
candidates from A have been checked, the exterior
orientation is improved by running a second bundle
adjustment including the new points.
After this first stage we have (one or more) groups of
three oriented images, some known object points and
some image points still to be numbered on the images.
To go further the operator looks for adjacent images
sharing at least three known object points and three
more targets with two of the oriented images. He just
assign the correspondencies, labelling the new points on
the new image. Based on the common set of points, the
above procedure is repeated and the new image is
incorporated in the block, while the number of targets
determined in object space increases. The procedure
continues adding new images but may eventually stop if
there are no images satisfying the above conditions. If
this is the case, the operator must find another group of
three images to restart with.
Figure 3 - The identification of the corresponding image points
521
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
