ich
stem
oto
ng
on
the
™
BEUTE
ct
O
at
CURRENT DEVELOPMENTS
It is significant that a growing number of research universities have
acquired this system and are pursuing independent, but complementary
approaches to the image matching/height determination problem, and to
other related tasks, such as digital orthophoto production and feature
analysis applications. Some of the other approaches to the image matching
problem being pursued are based on the least squares method, while others
are being based on concepts used in digital signal processing. Also being
investigated is the task of integrating the image correlation system with
a progressive sampling approach to point selection.
Our own efforts are being directed toward increasing the accuracy,
reliability, and speed of the current system, working within the framework
of the DEM grid collection by the VLL method. Faster processors and faster
video hardware are being considered. A current problem is what to do when
cases of poor correlation are encountered while stepping through the grid
of desired height points. It is believed that a solution should be sought
on two levels:
(1) Finding alternate sampling or computation strategies while the
instrument is still at the point in question, to try and resolve
the uncertainty while the image data is readily available, and
(2) Finding a more global, system strategy of ways to prevent such
occurrences in the first place. Determining how to efficiently
utilize the machine and the operator in case manual re-observation
is necessary, and deciding what data to record so that the point
may be automatically interpolated later from surrounding points.
With regard to item (1) it has been observed that on occasion the initial
window size selected fails to include sufficient gray-level information to
achieve good correlation. The system initially uses a small window of
approximately 9 by 9 pixels, and a small delta-z range of around 50
micrometers between the test levels (at photo scale). A sample image was
chosen from 1:30,000 scale photographs with very little detail at the
point to be measured. This small window yielded a very ambiguous
correlation function as shown in Figure 1. Enlarging the window size to
23 by 23 yields the correlation function shown in Figure 2. This curve
shows more of a central tendency but the shape is still insufficient to
determine a peak, or maximum point. Maintaining the 23 by 23 matrix but
increasing the pixel spacing to 150 micrameters yields the curve shown in
Figure 3. Utilizing a window matrix of 125(x) by 3(y) yields the curve
shown in Figure 4. Thus even though there may be insufficient image detail
at the point in question, there may be such detail (gray-level
modulations, edges, etc.) close by. A program should have the flexibility
to utilize such a dynamic windowing capability in order to obtain the
required levels of information necessary for image matching.
Another area from the same pair of photographs was selected in which the
information content was much higher. This yielded the curve shown in
Figure 5 for the 9 by 9 window, and the curve in Figure 6 for the
23 by 23 window. Both of these curves display a well defined peak in the
region of no parallax.
567