coordinates from 4,000 to approximately 100. The whole
measurement and filtering phase was rapid, taking
approximately 3 minutes per image, with software running on
an UItraSPARC, Model 140.
The image coordinates of the photo-control points and
reseau/fiducial marks were measured manually using the Erdas
Imagine remote sensing package. These measured image
coordinates were merged with the file of automatically
generated coordinates representing the floating marker points.
2.4 Automated data processing
The image coordinates of all photo-control and surface marker
points were transformed into photo-coordinates using a local
bi-linear transformation. This was achieved using a PC based
Visual Basic program utilising the measured and calibrated
positions of the reseau crosses, determined prior.
The filtering program used to reduce the number of measured
points generated a unique identifier for each measured point.
This was effectively a sequential integer determined by the
order in which the centre of gravity operator processed each
particle. Although such a unique point identifier is valuable, it
is of course essential that a surface target is identified using
the same number on other subsequent images. It was therefore
necessary to develop a program which would automate the
renumbering of points utilising two sets of image coordinates.
This problem is well known amongst photogrammetrists and
the traditional solution involves using epipolar geometry to
isolate the most likely matching candidate, (Dold & Maas,
1994). This approach requires knowledge of the exterior
orientation of the images, which could be obtained readily by
using the measured photo-coordinates of the photo-control
points. The main weakness with epipolar geometry is that the
epipolar condition only constrains the search for the valid point
along the locus of a line. The condition is most effective if
three or more cameras
are available, in which
case two or more
epipolar lines will
intersect. It is possible
also to constrain the ge
search along a section N
of the epipolar line by M6
enforcing some valid & :
search region in the re
object | space. Both \ :
constraints are often "3 \
applied, (Dold & Maas, S ac
1994). The approach N N
adopted 1n this situation \ N32
involved making both ^ X
direct and indirect use x4 \28
of the collinearity >
equations. Each photo- ^ X
coordinate on the left n
image and each photo- X
coordinate on the right Nd
image were used in a
simple algorithm to
initially compute an
; i . 2 metres
object coordinate. The
collinearity — equations
were then used in their Figure 2,
directions
102
Plan view...of 3D coordinates
direct form to determine photo-coordinates and subsequently
photo-coordinate residuals. These residuals were then summed
to provide a measure of matching quality for that particular
pair of photo-coordinates. This was then repeated for all other
photo-coordinates appearing on the right image and the
minimum summed residuals was judged to represent the
correct match. It was found useful to minimise the incidence of
false matches by implementing a spatial constraint in the
object space. This took the form of a 'band' of acceptable Z
coordinates and was implemented readily because object
coordinates were computed by the initial algorithm. Other
improvements which assisted the speed of the process included
the setting of a flag once a point had been successfully
identified which enabled subsequent data processing to be
skipped.
Once these data had been sorted and valid point identifiers
assigned to common points it was possible to derive object
space coordinates using a self-calibrating bundle adjustment.
Photo-coordinate observations to the control points were
included, their a priori standard deviations defined by the
'variation of coordinates estimation' (Section 2.1).
The derived physical centre of the targets did not represent the
position of the water surface due to the radius of the
polystyrene balls. There was also the additional, although
minor, systematic effect of buoyancy due to the weight of the
polyball. The combined effect of these two systematic errors
was compensated by subtracting a small offset distance to the
elevation of the computed three dimensional coordinates of the
floating net of targets. This was achieved using a spreadsheet
package.
2.5 Data quality
An important aspect of any surveying or photogrammetric task
is to assess the quality of the derived data. Estimates of
67
N
N
N
N
N66
62 A
x 58 \
\ N '59
65 » N
\ \ N
\ N98 s
60 N 52
N ‘51 \ 72
\
X7 “49 \ \
N \ \ \
\ \ 48 |
M \ \ \
«4 d j
s 43 \ 45
\ \ x '
\ \ 42
‘89 x \ \
\ \ \ \
A \ > 40
\
B6 ‘34 \ A
\ \ \ \
\ N X \ p
\
ne Ss s
\ gl X |
\ \ \ |
5 \ ^
YS een) ait aie on
£8
5 \ X \ \
\ 20 \ \ \
MN \ \ \
\
N \ e? \
N19 AT \ 23
\
N
\ \
N 15
\
\ 0
showing stream lines and
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
pre
cal
est
rar
pu
im
mc
de
the
pos
nol
ext
rev
the
Sys
flo
flo
mo
sur
2.6
Th
Sit
fur
sur
1S0;
su
