anbul 2004
ntention to
. The point
out on an
lytical with
€ the same
ie standard
10 um, the
the relative
ientation to
lation was
Papaharisi,
n, ISAT's
id of single
ced by the
zed by the
ng strategy
If this goal
t matches,
distributed
on Gruber
# of tie
points
1911
891
869
818
2199
ons
zd
AAT
id 5x5 von
of patches,
pattern an
; due to the
pattern the
0.047
0.055
0.079
0.018
0.148
0017
Ps
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
In Table 3, the values o, from the robust bundle adjustment
(both in pixels and in um) and the RMS values in object space
are presented. It may easily be concluded that:
e The standard deviation o, of the tie point coordinates
generally lies between 0.3 and 0.5 pixels or 4 to 8 um. This
result has been obtained although the expectation for the
accuracy of the image coordinates, as expressed in a priori
0, was set to 10pm.
e The o, value from the robust adjustment alone cannot be
considered as an indicator for the quality of the aerial
triangulation results. The reason is that in contrast to
analytical photogrammetry in AAT an appropriate point
distribution in each image and proper connections between
the images and strips are not necessarily ensured. (Heipke
& Eder, 1996).
Project | RMS on image | Mean Stand. Deviation (m)
X (um) | y (jum) X Y Z
1
2 2.9 3.5 0.028 | 0.029 0.045
3 3.5 33 0.020 | 0.021 0.035
4 4.1 4.1 0.050 | 0.050 0.084
5 3.6 4.5 0.021 0.021 0.040
6 56 5.6 0.076 | 0.076 0.131
7% 156.4 52.5 12.249 | 11.224 | 40.728
Table 4: Results of bundle adjustment for all points
* procedure stopped at 6" pyramid level
For scales 1:6000 and 1:8000 a small reduction of the accuracy
is noticed compared to the 1:3500 block. Obviously, the smaller
the image scale, the smaller the accuracy of the solution
obtained. For scale 1:8000 extra control points were added in
order to strengthen the solution, i.e. provide better geometrical
stability.
Project | Iterations | GCPs Max Residuals (m)
X y Z
1 11
2 10 11 0.062 (0.157. 0.101
3 7 11 0.081 | 0.163 | 0.097
4 7 8 0.085 | 0.125 | 0.120
5 8 18 0.120 | 0.104 | 0.144
6 12 11 0.149 |0.3157 0.292
7 10 0.043 | 0.060 | 0.030
Table 5: Residuals
The combination of scales 1:6000 and 1:8000 led to predictable
results, but important for the discussion. This procedure was
successful but it should be noted that a lot of attempts were
made to this direction. One of them was to increase the patch
window size to 120 pixels so that more pixels could take part in
the matching procedure. Moreover, control points were added
interactively. Therefore, this combination gave a solution with
impressive results (Table 5).
The combination of the scales 1:3500 and 1:8000 failed to
provide a solution. One possible reason was the drastic
difference between the two flying heights consequently the big
difference between the resolutions of the images (Figure 6).
Thus, the software split the block into two sub-blocks in order
to solve them separately. Without having additional known
DEM information, the block adjustment failed.
599
crue
he difference in resolution between scales 1:3500
and 1:8000 in the same level of zoom
ans: 4
Figure 6: T
3.3 Results of analytical solution
The results of the various analytical adjustments of the various
blocks or combinations thereof are summarized in Table 6.
Block Tie points (cm) GCPs (cm)
Oy Gy o, 9€. | 6, | 6,
1:3500 6.3 6.3 10.2 | 4.0 | 3.7 | 4.5
1:6000 5.1 5.8 10.9 | 4.7 | 5.1 | 6.3
1:8000 7.4 7:5 17.4 | 42 | 4.1 | 4.6
1:3500-1:6000 6.2 6.2 10.9 | 4.7 | 4.6 | 5.5
All scales 6.3 6.3 12.1 {46145158
Table 6: Results of the analytical adjustments
It is obvious that for the smaller scales the solution gives
relatively better results, mainly because the images include
more GCPs. On the other hand, larger scales contribute to the
solution with more accurate determination of elevations. With
the exception of the elevation accuracy of the 1:8000 images,
the other solutions are assessed as satisfactory. The better
accuracy in 1:6000 than in 1:3500 is probably due to the
appearance of more GCPs in a more favourable distribution.
The combined solution of the three scales proved to be a very
useful tool for the detection of systematic errors. Every single
solution usually presents small deviations, but the whole block
may depart from the correct absolute position within the
reference system. The combined adjustment with all images of
all scales, detects this as residuals in the measurements. The
large number of observations gives the capability for checking.
In addition, the unified solution seems to have values closer to
the average value of all and it presents uniform distribution of
errors. It presents no increase in accuracy but it is a more
reliable solution (it has more GCPs and measurements than the
others). The ellipses of errors have direction to the center of the
area and they present a big deviation outside the University
campus where there are few GCPs.
3.4 Conclusion
There is no doubt that the future belongs to the Automatic
Aerial Triangulation. It is common knowledge that the
initialization, point transfer and point measuring phases can be
successfully and reliably automated. However, geometric
stability may only be achieved through the increase of the
number of tie points. At the same time the distribution and the
number of GCPs is crucial, perhaps more than in the case of an
analytical solution. Hence the “replacement of intelligence with
redundancy” (Ackermann 1996) should be carried out with
extreme caution.
