382
To provide additional information, selected panels were imaged every minute over a half hour period whilst the humidity
in the control room was cycled. These data could then be used to asses, for example, rate of change in shape and to
evaluate any anomalies occurring with rapid humidity change. The images from each epoch constituted a data set of
175Mb, amounting to 4,378Mb over the three month duration of the project. These 10,375 images had to be measured,
photogrammetrically adjusted and analysed for deformation in 2,075 separate sets of adjustments.
3. DATA PROCESSING.
For convenience and to allow some means of checking image quality on site, data processing can be conveniently
divided into three discrete sections: on site image measurement and documentation; integrated target image matching
and bundle adjustment; and analysis and visualisation of deformation.
3.1 On site processing
On site processes included grabbing, consistent labelling and storage of images to disk using a PC based Windows
environment which was specifically developed for this project. Target location to sub-pixel accuracy was carried out
using a weighted centroid method (Clarke et al 1993). A minimal threshold value of 7 was selected to remove any
background information prior to computing target centroid locations. It was found that images could be stored and
measurements made of the several hundred targets in each image in the time that it took the conservator to reposition
each panel in the measurement easel. A complete measurement, including all 74 panels and two sets of calibration
data, could be obtained in about an hour and a quarter. At the end of the site process, data consisted of a series of
measurement files, numbered according to panel and viewpoint but with no reference as to target correspondence. The
images and image measurement data were compressed and transferred to City University for subsequent processing.
3.2 Target image matching and bundle adjustment
Target image matching was carried out using a multi-prpcessor SPARC 10 computer with 128MB RAM. Initial camera
orientations were obtained with reference to the scale bar targets, which were easily located on the images, and for
initial approximation purposes could be assumed to lie on a plane in the object space. A vector based camera
parameter approximation algorithm described by Fischler and Bolles was used for this purpose (Fischler 1981).
A 3D epiplanar match method (Chen et al 1993) was used to find corresponding targets in each set of image
measurements. This process was iterated with a bundle adjustment procedure whereby the camera parameters and
target coordinates were refined. The datum for the adjustment was based on inner constraints, but with scale defined by
the measured distances between the targets on each scale bar. To avoid blunders, possible with dense targets and
approximate camera orientation parameters required an initial constraint that each target appeared on all five images.
As the estimates of the exterior orientation parameters were refined, the constraint was lowered to four viewpoints, and
in the case of very deformed panels reduced to three viewpoints.
Whilst a simple counting based match method would probably have been more efficient, given the target regularity and
approximately planar objects in this specific case, the general method described was able to compute the location of all
panel targets. Difficulties were experienced with some of the very densely targeted panels simply because reprojections
of the imaged targets were indistinguishable within the noise level of the estimated camera orientation parameter
locations. Problems of this nature were solved by using a smaller object space reprojection tolerance such that fewer
targets were matched at each iteration. Some typical statistics from a panel bundle adjustment are detailed in Table 2.
Degrees of Freedom: Oo : No. Measurements
2802 0.337 2022
Co-ordinate axis X Y Z
Target RMS 66 um 63 um 85 um
Image RMS 0.54um 0.40um | --——-
residual (1/16 pixel) (1/18 pixel)
Table 2. Some parameters from the scale constrained adjustment computed for panel 1209 at epoch 0.
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995
Eac
a ci
sol
4.1
The
trar
pos
thre
ove
in €
viev
Fig.
An &
read
Profi
from
minu
auto
4.21
Fig. 8
IAP,