For matching two homologous image windows a common criterion of
maximisation of correlation coefficients is used in the SODAMS
system. The investigation of SODAMS was concentrated on analysis
of digital correlation and which strategies should be adopted in
this system. The main characters of SODAMS are summarized as
follows:
--Two-dimensional image windows are used, but only one-dimensional
shift i$ detected to determine xeparallax of conjugate point, be-
cause no y-parallax exists in epipolar line image.
--Hierarchical image data structure and the correlation procedure
from coarse to fine are only used for first several epipolar lines,
because at that time only little information (such as maximum par-
allax in current stereo-pair of photographs) could be referred and
it has to be considered to use a large search segment size.
--À prediction scheme has been designed in SODAMS to predict a next
point to be used to center a search area based on previously matched
points situated on preceding epipolar lines. The corresponding
search segment size is estimated simultaneously. Search segment
size relies heavily on the coarseness of the terrain surface repre-
sented by previously matched parallaxes.
--After the maximum correlation coefficient at a pixel is found,
a parabolic curve is fit over the three values on either size of
the maximum value to obtain the "real peak" of correlation coeffi-
cient function. Therefor, a subpixel measurement accuracy could be
reached, When the maximum correlation value is detected at either
extremity of the search segment, the algorithm will extend the seg-
ment size untill this occurrence has been avoided.
--When the gray-value standard deviation (variance) of reference
image window is less than a given threshold (low-contrast area),
the predicted parallax value can be applied directly. It means
that there is no need to do correlation in this case.
--Some factors for correlation quality controling have been com-
bined as one reliability factor to indicate this matched point is
either acceptable or ambiguous for further processing.
. value of the correlation coefficient
. distance of the correlation peak value apart from the
centre of search segment -- the predicted point
2.4 Detection and correction of blunders in correlation results
Even though some measures have been taken in the produre of corre-
lation, it is still hard to ensure all parallaxes measured by dig-
ital correlation to be corrent. Therefor, it is indispensable and
of great inportance to detect and correct the mismatching points.
During the process of correlation (2.3) a "on-line" debugging pro-
dure for ambiguous matching points has been carried out simulta-
neously. As a matter of fact the prediction has tacken unreliabi-
lity of correlation into account. But a post-processing scheme for
further dection of blunders is applied as an independent component
in SODAMS, which consists of two parts:
- 677 -
