clearance. Obviously in this case such vectors are not
useful, except to demonstrate that control points (9,10,11)
appear stable and that point (14) on the supporting jig has
moved in a similar direction to the targets on the blade.
5. DATA PROCESSING
Initial orientation parameters for each of the cameras
were computed by the method of Fischler and Bolles
using the locations of four of the surveyed retro-target
control points (Fischler and Foley 1981). The tracking
procedures used to automatically identify common targets
between similar images at each epoch meant that the
target correspondence problem was significantly reduced.
Targets imaged at each of the five camera stations in
epoch 10 were matched using a target correspondence
algorithm developed at City University (Chen 1995). The
method is founded on a three dimensional epiplanar
geometry, iteratively incorporated within the bundle
adjustment procedure. To avoid blunders, which can
occur with dense targets and approximate camera
orientation parameters, initial constraints were added to
ensure that each target matched appeared on all images.
As the estimates of the exterior orientation parameters
were refined, the constraint was lowered to four
viewpoints, and finally reduced to three viewpoints to
collect any remaining unmatched targets. In this case the
method correctly identified and matched all of the targets
in five iterations. A common numbering scheme between
all epochs could then be simply applied by cross
referencing all target measurements within the
appropriate images in the primary epoch. Any ambiguities
caused by excessive target movement or occlusion
during the course of the experiment were resolved by
reapplying the epiplanar method to any remaining
unlabelled target points.
During the experiment, the blade was adjusted about a
pivot point so deformation could not be estimated with
respect to the survey datum. Instead, deformation
analysis techniques were centred on a ‘free network’
bundle adjustment. To commence an analysis of the
movements occuring in the rotor blade, the photoco-
ordinate measurements and target co-ordinate estimates
of targets on the rotor blade at epochs 10 and 11 were
isolated. An initial common datum was defined by using
the epoch 10 target co-ordinates as starting values in an
iterative least squares estimation based on the method of
‘inner constraints’ (Cooper, 1987). Since these starting
values were final co-ordinate estimates based on the
survey datum, scale was preserved.
A global congruency test (Setan 1995) (using a = 0.05) to
determine whether any significant movement had
occurred between the two epochs was computed by
analysis of the target co-ordinates and their associated
covariance matrices. The global test indicated significant
movement, so it was necessary to locate the unstable
points. These were removed one at a time because the
datum had to be re-defined with respect to the remaining
points. A localisation procedure modified from Fraser and
Gruendig (1988) was used for this purpose. The
procedure consists of decomposition, re-ordering and S-
transformations (Baarda 1973; Strang van Hees 1982)
until the congruency test (using a=0.01) (Cooper 1987)
was accepted. In this manner a group of stable points
was identified. The displacements and stochastic
attributes of the remaining points relative to the new
datum were then derived and verified. This process was
repeated as required to allow a comparison to be made
between pairs of data from different epochs.
Some parameters from the adjustment computed with the
‘inner constraints’ epoch 10 data are shown in table 3.
The RMS image co-ordinate standard deviations are
derived from the mean of the measurements made from
each set of three images. Since the target image quality
was relatively poor, the frame averaging algorithm set a
global image measurement standard deviation of 1/5th
pixel.
Table 3 Some parameters from the epoch 10 adjustment
Degrees of 6.2. 0.220 | Measurements Targets
Freedom: 333 | © 542 62
Axis X Y Z
Target RMS 0.09 mm 0.13mm 0.07 mm
(* og)
Image RMS 0.73 um 038um [| --——
residual (1/11 pixel) (1/23 pixel)
496
Oo = standard error of an observation of unit weight.
The relatively low object space precision, 1:30,000 of the
object space, is a function of network limitations imposed
by the number of cameras which could be used in the
system and object space restrictions on their location and
orientation.
Computing a bundle adjustment including observations of
straight lines at different distances gave a significantly
larger variance factor, of the order of 1. This effect was
isolated within the bundle adjustment to large residuals
on the straight line observations caused by primarily by
noisy line image observations. Calibration was instead
carried out by computing radial and tangential lens
distortion parameters from the mid-distance set of straight
line images and holding these values fixed within the
adjustment. Further work to include complete straight line
calibration data, initially under controlled laboratory
conditions, will be carried out as the technique could
prove invaluable given the very large distortions present
in low cost 'C' mount lenses used.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
©
Figure :
ellipsoic
Comput
confider
Integra
develop
of 10x e
and 11
during
differen
significe
significe
the bott
physica
matherr
procedL
and 17,
blade. /
respect
Figure
