,y and z. This
for the imagery
but featureless
ject some form
relation. The
1ares matching
ed and tested -
sensing image,
these, the best
is found to be
ze (Figure 3).
Kodak DCS200
ct distance was
1:1.5 and the
: approximately
xel size on the
use convergent
o distance ratio
thus a more
images Were
id converted to
> brightness and
s also used to
control points.
The back was also photographed with an Officine-Galileo
metric camera and a DEM of the back with a 5 mm grid
interval was subsequently measured from this imagery in the
Zeiss P3 analytical plotter. This provided reliable reference
data for the accuracy tests, the standard deviations at the
control points being +/-0.2 mm in x and y and +/-0.4 mm in z.
This DEM was taken to represent the true shape of the back
when assessing the accuracy of the digital method.
Measurement of the Kodak DCS200 imagery was undertaken
in the R-Wel DMS within the Softcopy Photo Mapper module.
This module allows the orientation of a stereomodel through a
separate space resection of each photograph using a minimum
of 5 control points. The 4 corner points (the principal point
was assumed to be in the centre of the image as defined by
these points) and 6 control points were measured on each
image and the results of the orientation are shown in Table 1.
correlation matrix. For this particular work, the generation
times were 2, 10 and 27 minutes for 15, 10 and 5 post spacings
respectively for a 17x17 correlation matrix, with the 11x11
matrix taking 0.5, 4.5 and 14 minutes respectively.
The resulting DEM can be viewed in the ‘View 3D
perspective’ module as a wireframe model from any desired
angle and with the image draped over it if required. It is
possible to edit erroneous points either stereoscopically or by
running a smoothing filter over the DEM. The models were
not, however, edited and were simply rectified, then translated
to xyz format for export to the LSS software package. At the
rectification stage, the DEM can be clipped to the required
area. It is also possible to mask areas where the correlation
fails, such as the edges of the mannequin in this case, but
again, this facility was not used.
Orientation Image 1 Image 2
parameters
X (mm) 1571.213 2388.558
Y (mm) 9680.940 9674.683
Z (mm) 1910.263 2011.815
© (degrees) 2.4 2.4
$ (degrees) -21.1 12.1
x (degrees) 0.5 -1.2
pixel size (mm) 0.410 0.415
| image scale 1:447 1:452
RMSE (pixels) 1.45 0.62
RMSE (mm) 0.59 0.26
Table 1 - Orientation parameters for the images used.
After orientation, the area of the stereomodel was defined and
the resulting stereo images were used for the generation of the
DEM's. The number of pixels between the DEM points (post
spacing) and the dimension of the correlation matrix for the
matching have to be specified, as do the maximum and
minimum z values expected to be encountered during the
creation of the DEM. In this work, DEM's were generated
using post spacings of 5, 10 and 15 pixels and with correlation
matrices of 11x11 and 17x17 pixels, to determine which gave
the most accurate results. Inputting the elevation range is an
iterative process. The expected elevation range was input and
the software gave the actual elevation range after measurement.
The DEM then has to be regenerated based on this new range
and the process repeated until the output range falls within the
input values. The input elevation range definitely affects the
resultant DEM and further investigation into this is underway.
For each generated DEM, the R-Wel DMS indicates the
completeness of the correlation and this was normally in the
order of 99%. This value is merely an indication of the success
of the matching and gives no indication of its accuracy. The
time taken to generate the DEM's obviously increases with a
smaller post spacing (the 5 post spacing DEM had 41475
points compared to 4582 for the 15 post spacing) and larger
409
Figure 4 - Wireframe model of generated DEM of the back.
In order to make the comparison in LSS, a survey needed to be
set up and the data from the Zeiss P3 read in. A separate
survey then had to be created for each generated DEM. The
DEM measured in the Zeiss P3 was used as the base survey
and each generated DEM was subtracted from this to create
differences between the two data sets at around 2000 points.
The integrity of the original Zeiss P3 data was thus maintained.
The differences were then analysed by reading the LSS files
into a spreadsheet and calculating the standard deviation of
these differences. The results are given in Figure 5.
4.2 Discussion
Examination of Figure 5 indicates that the best results were
achieved with the densest post spacing and with the largest
correlation matrix - in this case, a post spacing of 5 pixels and
a 17x17 correlation matrix. With this configuration, the
standard deviation of the depth differences over the DEM was
0.94 mm which is equivalent to 2.3 times the pixel size. A
standard deviation of 1 mm is considered adequate for a
number of anticipated applications that could arise through the
use of a low cost, off-the-shelf system such as this and so the
10 pixel post spacing (standard deviation 1.02 mm) may also
be acceptable.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
