makes use of Hamilton’s Quaternion representation for
rotations.
Experience has shown that this solution may not be very
accurate: a least squares solution based on collinearity equations
is therefore computed to improve it; at this second stage also
other GCPs available in the image may be included. To this
aim, the 7 parameters determined by RESECT are used to
linearize equations (1). In a previous paper (Albertella and
Scaioni, 1999), we compared the collinearity equations in
general form (1) to that in normal form (where the scale factor
has been eliminated): we found that the former have a larger
convergence radius while the latter allow a more accurate
computation of the solution. We therefore use both in that
sequence, to ensure that the space intersection will provide good
approximations for tie points.
The analysis of singularities in the computation of the
parameters or the conditions for the uniqueness of the solution
have not been investigated analytically; a general discussion on
the topic may be again found in Wrobel (2001). Some empirical
tests with simulated data has been performed to gauge the
sensitivity of the algorithm.
In this respect, a problem coming from the original version of
RESECT was the selection of the three points among the four to
be used to write down equation (4): it turned out in fact that
different combinations of them may give rise to different
solutions. We got rid of this ambiguity by simply computing the
solution for all possible combinations of the set of four points
and keeping that with the smallest RMS of the residuals on the
image coordinates, if smaller than a given threshold. From the
simulations we performed, this always turned out to be the
correct one.
If more than four known object points are available, those
defining in image space the quadrilateral with the maximum
perimeter lenght are tried first; if no solution satisfying the
above mentioned threshold can be found, all combinations of
four points are explored.
Figure 1 — Geometric meaning of symbols used in RESECT.
Computation of Approximate Ground Coordinates
The computation of a least squares bundle adjustment requires
also the approximate object coordinates of tie points. Neglecting
the proper definition of the stochastic model, it is
straightforward to compute them by writing down the
collinearity equation in linear form; nevertheless, our
experience suggests that results are not accurate enough when
dealing with convergent images. Then we implemented an
approach based on a vector representation of two homologous
rays, presented by Cooper and Robson (1996). The formulas to
determine the object coordinates of the point are given in the
following, according to the geometric scheme in Figure 2.
Given a pair of oriented and calibrated images, let u and v be
two unit vectors projecting two corresponding image points in
object space. By properly choosing two positive scalars 7 and Ç
the intersection of vectors 7u and Cv gives the point A in object
space. Due to errors in image coordinates and in orientation
parameters, we will have a residual parallax vector, given by:
p=-uth+y (7)
where b = [1 b,/b, b./b,] is the baseline vector. Parameters N
and Ç can be determined so that length of p is minimum:
.-(u: v)(v : b) * (ub)
1-(u-v)
_(u-v)(u-b)—(v-b)
7]
8 1-(u-v)?
(8)
0
Let 7%, Ç be the solution of the minimum problem. The
intersection point A will be located in the midpoint of p:
a -— ru * 5 p — ^ (ru * b + wv) (9)
while the magnitude of the parallax vector can be assumed as an
indicator of the accuracy of the position of A:
Po =-70u +b + Gov (10)
This procedure yields straightforward the result and does not
need further iterations.
Figure 2 — Geometric scheme of space intersection.
TESTS WITH SIMULATED AND REAL BLOCKS
The check of the procedure described so far has involved
several photogrammetric blocks, covering both terrestrial and
aerial environments, different kind of sensors (metric analogue
cameras and calibrated digital cameras) and geometry of camera
stations. Some tests have been carried out also on simulated
blocks. In this section, we refer to one simulated example and
two real blocks, deemed representative of two different block
geometry often met in practical applications.
—456—
At first,
order t
configui
points, ¢
a series
gravity |
camera
of a gric
axis poi
coordin:
paramet
test has
whether
to the |
solved.
sense)
critical
recover:
for the «
The sec
procedu
blocks.
Wiki
Blocks
first the
the sec
and les:
the foll
The fir
images
sculptu:
regular]
at least
block h
a small
have be
by this
comput
the rol
RESEC
into the
increas
their ac
every b
constra
recomp
into the
determi
instanc
but ime
strengh
The sec
Cloiste:
feature
photos
taken L
unfortu
acquisi
the ima