the
ni-
and
on-
vith
ins-
idi-
-tO-
(2))
cifi-
ited
the
ins-
(3)
fac-
dis-
for
s. It
the
nge
o 3
ion
ach
ter.
at a
jeld
n
)36
)31
25
)18
ing
eir
37)
rge
t of
vity
sed
Ihe
[he
tor
Ses
ion
vith
im-
5 o ô
Version [AP| Co| Ch r Og % Y Z ux us uz x Hy
| [mm] | [mm] | [mm] | [mm] | [mm] | [mm] | [um] | [ium]
1 0| 30| 56| 11554| 11.7| 0.027| 0.023| 0.035| 4.207| 4.063| 3.657| 12.13| 11.98
2 10| 30| 56| 11547| 0.36| 0.027| 0.023| 0.035| 0.086| 0.058| 0.114] 024| 0.22
3 10| 30| 56| 11547| 0.42| 0.023| 0.019| 0.029| 0.068| 0.055| 0.084| 0.19| 0.17
4 10| 30| 40| 9879| 0.42| 0.023| 0.019| 0.028] 0.062| 0.039| 0.056| 0.16| 0.14
5 10] 30] 38 8969| 0.38] 0.026| 0.0201 0.032] 0.052] 0.040] 0.048] 0.13] 0.13
6 10] 30{ 38 8963| 0.46] 0.033] 0.025] 0.039] 0056| 0.039 0.061] 0.14] 0.14
Improvement 1 1 2 32.5 1.0 1.0 1.0 48.9 70.1 3211 505] > 545
Improvement 2 | 3 0.9 1.2 1.2 1.2 1.3 F2 1.4 1.3 1.3
Improvement 3 | 4 1.0 1.0 1.0 10 1.1 14 15 12 1.2
Improvement 4 | 5 Ll 0.9 1.0 0.9 1.2 1.0 12 1.2 14
Improvement 6 | 5 12 1.3 1.3 12 1.1 1.0 13; 1.08 1.08
Table 2 Results of bundle adjustment and accuracy
= A Scale for Points 500.000 [mm]
verification. Scale for Vectors — 50.000 [micron]
AD... rest Number of additional parameters p
C9 iino deni. Number of control points = d x €
Ch ee Number of check points Targets on rods of front structure
T axccsessapecisssees Redundancy 7
genns: Variance of unit weight a posteriori
Ox, Oy, Oz . Theoretical precision of check point coor-
dinates X
Ux» Lys Hz..... Root Mean Square Error from comparison
to check point coordinates in object space
Hays Hy coeecnencs Root Mean Square Error from comparison
to check point coordinates in image space
The plots of checkpoint residuals of version 4 exhibit
further systematic discrepancies which were traced. The
small imaging scale and oblique imaging of targets on
the rods of the testfield structure led to an influence of the
(darker) background on the target area. This resulted in
displacements comparable to those of local illumination
gradients. Targets which were imaged too obliquely were
eliminated for version 5. The relative precision of ver-
sion 5 corresponds to 1 part in 100000 for both the X and
Y axis. The relative accuracy is 1 part in 50000 for the
same axes. The accuracy in image space corresponds to
1/85 of the pixel spacing. An accuracy of 1/50" of the
pixel spacing can be attained when using a minimum da-
tum only. This difference might appear large, but must be
partially attributed to residual effects of local illumina-
tion gradients, which are partially absorbed when using
more control points. These effects reached 0.03 pixel for
the targets which were eliminated in the step from ver-
sion 3 to 4 and are assumed to reach 0.02 pixel for other
targets still in the data set. Systematic discrepancies can
also be detected for targets in columns as shown in the
plot of the differences to checkpoints in Figure 15. The
overall accuracy improvement in image space attained
by using a smaller template and eliminating targets with
large degradations (version 2 to 5) is a factor of 1.8 and
1.7 in x and y.
& * S. o ^
Targets in columns on wall
Figure 15 Differences to check points of version 5 (test-
field as viewed from top).
3.4 Accuracy with PLL Line-Synchronization
The imagery was acquired at identical positions with pix-
elsynchronous sampling and PLL line-synchronization
within 30 seconds of each other. It is thus possible to
compare the effect of the synchronization on the three-di-
mensional accuracy under practically identical condi-
tions. The image coordinates were again measured with
LSM using a 5 x 5 template and the targets which were
eliminated due to local illumination gradients and too ob-
lique imaging conditions on the rods were removed from
the data. Figure 16 shows a comparison of the image co-
ordinates for 6 frames acquired at the left topmost station
(see Figure 12). The average displacement between the
two images of 0.125 and 1.988 pixel in x and y was re-
moved before the plot. The average displacement in x
was attributed to a difference in the sampling point and
the difference in y was due to differences in the settings
of the frame grabber for the selection of the active image
region in vertical direction for the two acquisition modes
(the determined value is only 0.002 pixel off the actual
difference of 2 pixels). The large geometric deformation
and instability (especially in x) of PLL line-synchroniza-
tion is apparent from the figure. The RMS value of the
