image is then usually an artificial and ideal version
of the signal to be located in the other.
The matching was done off-line and interactively
using a window-based interface to the algorithm.
This allows all algorithm parameters to be set to
optimum values for a particular matching problem,
plus assessment of the result by visual inspection.
The program transforms matched pixel coordinates
directly to the photograph frame if the appropriate
level of transformation exists.
Different templates were used for the fiducial
marks and the signalised points. Both were
approximations of Gaussian signals and differed
only in size. Consequently, different patch sizes
were also used: 7x7 and 5x5 respectively, such
that no background signal was used in the
matching. One set of parameters for the matching
was selected empirically and thereafter the aim was
to use the chosen set in as many cases as possible.
For each photograph, the fiducial marks were first
matched and the inner orientation determined as an
affine transformation. The fiducial mark signals
are usually well defined and the matching can be
done without problem. In one or two cases, dirt on
the edge of the signal disturbed the solution, but it
was possible to overcome this by selection of a
different parameter set.
The signalised points were not so straightforward
to match; the images displayed a great variety in
size, shape and contrast. On the average, the points
occupied a circle of 3 pixels diameter (original size
on the ground 40 cm). Quality control in this
investigation was done by visual inspection: if a
point could not be matched satisfactorily with the
chosen parameters, then an alternative optimum set
was found - usually a smaller patch size or
restricted image shaping transformation. Of the
279 points, 74 (27%) were so matched. The final
solution was assigned a quality rating between 1
(bad) and 5 (good), which related only to the bias
of the solution from the point centre. These classes
were reduced to three for the final analysis. For
each point, the various statistics resulting from the
least squares matching were stored and so a
comparison was possible. The results can be
summarised as follows:
Table 1. LSM statistics mean values by class
Class
# of
points
Sigma
<?o
Standard Devns
a x (pel) <7 y (pel)
Corr’n
coeff
1/2 Bad
19
16.8
0.112
0.121
0.90
3 OK
69
14.1
0.081
0.082
0.94
4/5 Good
191
12.2
0.064
0.064
0.96
Although an expected improvement in the values
does exist, the differences are not distinct. When
the distribution of the values is also inspected, it is
found that some points classed as 4/5 have values
characteristic of class 1/2, and vice-versa. The
problem lies in the main with asymmetric signals,
where the final matching solution will inevitably be
pulled away from the visual optimum. The bias
cannot be directly determined from the matching
statistics. Asymmetry was found to be due to three
sources: noise in the photography (eg dirt); shape
and layout of the signal, particularly if the location
stripes were too close to the target (see example
images); or perspective distortion. The latter of
these was not the most significant. It should be
stressed that no point could not be matched.
Complete failures can usually be trapped, using
characteristics such as many iterations or unusually
low correlation coefficients, for example. Due to
the interaction in the matching, no points were so
rejected; the problem here concerns the quality of
successful matches, in the sense that matching was
possible.
Figure 3. Example signalised point images: on
the left a good one, on the right a bad.
4.5 Bundle Adjustment of the Observations
4.5.1 The Datasets and Configurations
The results of bundle adjustment on two datasets
will be presented: the manual observations and the
digital observations including all matched points.
Both datasets had no distortion corrections applied.
Two control point configurations were selected:
“PPO” with all points as full control points; and
“PP2” with 8 full control points around the block
perimeter and 4 height points in the centre (see
figures 4a and 4b). The block had a certain
asymmetry: the larger size of the original meant
that control points on the eastern side of the block
were in general only visible in one strip, ie had only
two rays. On the opposite side, the perspective
centres formed the control point boundary and
therefore most points existed in two strips.
Additional parameters were used in the
adjustments. Versions with 0, 12 and 44 were
computed, using the program Bund of the ETH. In
all cases, a priori values of 2.5 pm for the image
points, and 1 cm in plan, 2 cm in height for the
ground control points were used.