Operator 0 3a 3b
Least squares SD SD R SD R
A0 (0.01 pixel) 11.5 9.9 «0.05 12.5 «0.27
A1 (0.01) 2.41 1.7: .-0.06 2.2.5. «0.26
A2-(" 2g 1.9 1.7. 40.31 2.0'* -0.26
HO (grey value) : 2.0 0.7 2.9
H1 (0.01) : 6.4 1.9 8.8
163 üpergtar.id uns. epa Quabr o men Bin dt SOAR umb
Cross correlation : CC CC R CC R
mean maximum 0.72 0.96 +0,57 0.44 40.41
The results show that the precision of geometric parameters is low for
both methods, compared to the previous interest functions.
For A0 and A1 in the maximized case (3a), the rank correlation R shows
no dependency on the interest value. Further the precision in shear
A2 decreases with increasing IV.
In the minimized case, all geometric unknowns decrease in precision,
when the interest value reaches its local extreme.
Accordingly, the comparison of products is not a suitable operator for
increasing the geometric precision of the least squares matching.
The overall best radiometric fit is obtained using variant 3a, and in
agreement with this, the worst is obtained by 3b.
The same result is obtained for coefficients of cross correlation.
Although a significant similarity is obtained using 3a, this has no
relevance for determining parallaxes and slopes. For instance,
homogeneous areas in the photographs have high auto- and cross
correlations, but provide no structures for determination of parallax.
4. CONCLUSIONS
Using point densification ensures reliable height determination.
Cross correlation provides good convergence in the adjustment.
Precision in point determination can be optimized using interest
operators. The following improvements were obtained:
- parallax almost 40 %, slope between 30 and 40 %.
- With the exception of auto-correlation, the differences between
operators were relatively small. The best results were obtained
with grey level gradients in the x-direction.
Suggestions for further research:
1. Determination of threshold levels, which enable early rejection
of features and the possibility of examining several features within
a search area.
2. Two-dimensional search for suitable points. This leads to an
interest function IV(x,y) and the creation of a completely irregular
grid.
3. Recognition of discontinuities followed by correlation using
edge detection instead of the grey level methods described above.
- 571 =
