The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
931
4.4 Parameter Correlation 5. INDEPENDENT ASSESSMENT
Considering again case 3, the largest correlation coefficients
existed between the perspective centre co-ordinates (PCCs) and
the principal distance (maximum magnitude: 0.83) and between
the PCCs and the rangefinder offset (0.67). The maximum
correlation magnitude between the rotation angles co and <j) and
the principal point was quite favourable at 0.63, though it
should be recalled that the 10 model did not include decentring
distortion terms, which were found to be insignificant. The only
noteworthy correlation between the 10 parameters was 0.71
between c and do- For the rest of the range APs the maximum
correlation with any other parameter (EO or IO) was 0.20.
Addition of the empirical terms in case 4 changed the situation,
with maximum coefficients of 0.90 between e 2 and en, and 0.64
between do and e 4 . In both cases 3 and 4, though, all APs were
statistically significant in terms of the ratio of the estimate to
the standard deviation.
4.5 Scale Error Estimation
Addition of the scale error, d l5 to the case 3 adjustment did not
significantly change the residual RMS measures. A high
correlation of 0.86 existed with do but, interestingly and
encouragingly, the correlation with c was only 0.44. These
could be reduced by increasing the depth variation in the target
field. Though statistically significant, d t was excluded from the
final AP models due to its lack of impact on the other
performance measures. What this demonstrates is that the scale
error parameter can be successfully estimated by including
easily-measured spatial distance observations in the self
calibration adjustment, though it is conceded that a large
number (33) were used in this experiment.
4.6 Lens Distortion Modelling Results
The estimated radial lens distortion profile, 8r, of the SR-3000
is plotted in Figure 8. Though the amount of distortion is quite
high, -205 pm at the maximum observed radial distance of 4.55
mm, this represents only about 5 pixels due to the 40 pm pixel
size. The corresponding Ict error envelope is not perceivable
since the estimated standard deviation of ki was almost two
orders of magnitude (76 times) smaller than k] itself. The higher
order terms of the radial lens distortion model were found to be
insignificant.
5.1 Accuracy Assessment
To assess the accuracy improvement gained by the different IO
model cases, the object space co-ordinates of the surveyed
points were compared with those determined from the LRC. A
rigid body transformation of the LRC-determined co-ordinates
onto the surveyed co-ordinates was required for each of the 6
independent images. The overall RMSs of co-ordinate
differences calculated from the total available set of 188 points
are given in Table 3. The case 2 IO set yields improvement in
only one dimension (X) and the accuracy actually degrades
slightly in Y and Z. Clearly there is considerable improvement
realised by using the physical AP model (case 3). In case 4
there is more overall improvement in Z relative to case 3, but
there also was a slight degradation in X, which may be due to
over-parameterisation with the empirical terms.
Case
RMS
% improvement
X
(mm)
Y
(mm)
Z
(mm)
X
Y
Z
1
47
57
70
-
-
-
2
42
63
74
11
-11
-7
3
31
14
43
34
75
37
4
35
15
36
27
74
48
Table 3. RMS of check point differences and % improvements.
-500 . -500
Y (mm) X (mm)
a)
Y (mm)
b)
Figure 9. Uncorrected and corrected planar target data, a)
Figure 8. Estimated radial lens distortion profile. Isometric view b) side view