The precision of parallax AO increases with both operators by 1/3.
For both operators the dependence is the same, -0.61. As the interest
value increases, the standard deviation decreases (hence negative
dependence).
The improvement in precision of tilt Al (= slope in x-direction) is
less, as is the dependence on the interest value.
Shear A2 (= slope in y) is clearly better defined, as R indicates.
This is to be expected as edges in the y-direction are sought which
will easily show shear,” but little difference: for tilt -in the
x-direction.
Dependence for radiometric transformation has not been computed as
this is not important for DTM-generation. However, increase in
precision can clearly be seen, especially in variant 1b.
The coefficients of cross correlation also show improved similarity
using the operators, although dependence on interest value is not so
high.
The main advantage of the 2nd variant lies in the improved radiometric
fit. However, as this is not important for the DTM and involves more
computational effort, method 1a is preferred.
3.2 Differences
This is used to test more general shapes and is based on the
comparison of neighbouring pixels throughout the window.
a) Summation of differences of grey levels (in x-direction only)
IV 3 ?.l g(x1,y) - g(x2,y) | , Where x2
1 7] i
Hu:
=
—
Ca.
Hu
—
=
><
If pixel pairs are chosen with a separation between the pixels
(x2 + 2 ..) the result responds to a low pass filtering of the
window. However, this had little effect on the results.
b) Summation of squared differences
IV -2:2.( g(x1,y) - g(x2,y) 2 , where x2
14)
nu
x
+
| 5 4:.MY4j uj aod. MX
This is a special case of a Moravec-operator, which compares pixel
pairs in all directions, see also /5/. Again pixel pairs with
greater separation can be used.
c) Summation of second differences
11-2910 -02} , where i21... M, jd uu MX
pj
and D1 = | g(x-1,y) - g(x,y) | , 025196.) - 9691.) |
With this operator the curvature of the grey level function in the
x-direction can be examined.
- 569 -
