merging process is performed where close neighbouring
object points are merged into a single point, with
appropriate re-labelling of image point observations.
e Final bundle adjustment: A final bundle adjustment with
sensor self-calibration is carried out for all validated
observational data.
The final computation in three stages ensures robust and
accurate results, without the influence of observational blunders
or gross errors made in automatic feature matching and
labelling. Throughout the ‘fast’ bundle adjustment, the
performance of the process is successively improved with each
iteration, especially within VM networks comprising large
numbers of images and points.
6. PERFORMANCE ANALYSIS USING PRACTICAL
MEASUREMENT PROJECTS
The performance and reliability of the presented algorithms, as
implemented in Australis, have been proven in numerous
practical applications covering networks of various size and
objects of various scale. Here we use two practical measurement
projects to illustrate the performance of the automated VM
procedure developed. The projects where conducted with two
different cameras, and the two objects were of approximately
the same size, the first, Project 1, being a car door with 130
retro-reflective targets (Figure 8), and the second, Project 2, a
plastic mannequin with more than 300 targets (Figure 9).
Whereas the car door was recorded with six images using a GS7
INCA camera, nine images were recorded with a Kodak DCS
420 camera in the second test project.
Figure 8. Image from car door project, Project 1.
For quality assessment of the automated measurement process,
the correctness and accuracy of the results were analysed. It was
also investigated if all possible target information was used for
the final bundle adjustments. Additionally the performance of
the process on two different computer systems was assessed. An
initial quality statement can be made by examining the RMS
value of the image coordinate residuals from the final bundle
adjustment.
In the car door network, an RMS value of 0.15 pm (better than
0.02 pixel) was reached, which is an appropriate accuracy for an
INCA camera project. Though the plastic body was imaged in a
9-station network, the final RMS of the residuals reached only
0.24 pm (0.03 pixel). The DCS 420 camera did not have a
stabilised interior orientation, and consequently the achieved
RMS is also assessed as an acceptable value under the
circumstances (the same camera often yields residuals at the
0.25um level in Australis projects with retro-reflective targets).
The focus of attention in these test measurements was upon
internal rather than external accuracy, though it is noteworthy
that in each project the mean standard error of XYZ object point
coordinates surpassed a relative accuracy of 1:100,000 of the
object size.
The investigation of the two final VM networks showed that all
targets were both found and computed correctly in each project,
and no additional non-targets were triangulated. Analysis of the
imaging-ray statistics for object points showed that most targets
were measured in all images, though occlusions or highly
oblique incidence angles for certain targets prevented this in
some images. The results of the automated process were very
satisfactory and it should be mentioned that the data cleansing
procedure within the 3-stage bundle triangulation process
improved the results of the overall measurement, especially for
the second, 300-point project.
Figure 9. Targetted plastic mannequin, Project 2.
Regarding performance, the projects were computed both on a
Pentium II Celeron 433 Mhz computer with 128 MB RAM and
on an AMD Athlon XP 1600 computer with 512MB RAM. The
Pentium II system took 12 and 30 seconds, respectively, for the
complete measurement process for Projects 1 and 2. The
corresponding computation times with the AMD system were
substantially shorter, being 7 and 14 seconds. Details can be
found in Table 1. These computation times are certainly
acceptable for practical automated off-line VM, and it is
doubtless the case that improvements in PC technology will
bring even faster processing times.
Table 1: Computation times for the automated measurement processes in Australis (all values are in seconds).
PIOCeSSOT Project Scanning and | Correspondence Fast Bundle Data Final Bundle Overall
EO detection Process Adjustment Cleaning Adjustment Process
AMD Athlon XP 1 3.9 0.3 0.5 0.2 0.7 13
1600 2 5.4 1.4 2.1 0.4 2:9 14.3
Intel Pentium IT 1 9.0 0.5 1.0 0.2 1.5 12.1
Celeron 433 Mhz 2 12.4 4.5 $3 0.4 72 29.9
—66-—
v3
V d ri kA CR» Ure Ces rA ei pp
a f^^ E CR
pte
]
"zh = "x3
