sach
Irger
Il be
seau
f the
on a
ntral
; cut
| out
hich
| the
| the
four
has
Tie
| the
reas
one
sach
d as
ofa
r to
1the
id is
ixel.
Il be
ister
rest
than
alue
IPs
isa
ition
f the
their
iced
hest
ions
) not
evel,
t we
ly in
| be
are
e by
ning
n as
t we
ance
in gray value is selected as the final aerial triangulation
point (TP).
After the TPs in the master tie image have been
determined, their corresponding approximate locations in
the slave tie image must be found for serving as initial
starting positions for the fine LSM matching. The search for
their corresponding locations in the slave tie image is also
done through the image pyramid. First, the location of the
TP just found in the master tie image will be traced up to the
highest level, from there the corresponding location in the
same level of the slave tie image will be determined by
finding the maximal Normalized Cross Correlation (NCC).
Than these locations will be traced down through each level
of the slave tie image by finding the maximal NCC to the
corresponding master window in the same level. Since the
tracing is done from one level down to the other, which
means the initial position in each level is fairly accurate, the
search window needs not be very large, except for the
highest top level where we have no information for the initial
position and a blind search in the whole image is
necessary.
While going through the pyramid levels, a threshold of 0.65
is set for the NCC value in order to avoid wrong matching.
Whenever this threshold can not be passed, that TP will be
abandoned. If at the end there is less than two points left,
the system will issue a failure message and switch to the
manual selection of the interest windows both in the master
and in the slave images. The matched position in the lowest
level ofthe slave image will be considered as the conjugate
position to the TP in the master image. Its coordinates go
into the next stage of fine matching. The searching
procedure can be illustrated by Fig.1. The arrows indicate
the direction of tracing.
5. Matching Strategy
Based on the initial matching position obtained in the
previous step, the fine LSM can begin. But by trying to find
a proper window size for the LSM, we have found out that
there is no such thing as the best window size. Different
window size will give different converging position (or no
convergence at all). Some times when one window size
failed to converge, by changing the window size it could
come to a convergence. Therefore we gave up the effort to
find a best window size for all cases. Instead, we let each
point to be matched by several different window sizes and
try to find the best result from them.
We know that unless the LSM diverges, it will always give a
result (the two shifting parameters between the windows).
Although the posterior sigma naught reveals some
knowledge about the quality, but its power for judging the
correctness of the result is too weak. Therefore we use the
NCC value as a check. Prior to the begin of the LSM, there
was a NCC value for that initial position as computed in the
399
previous step. After the LSM converges, new NCC value
can be calculated between the target window and the
resampled matched search window. The relationship
between the master target window and the slave search
window before and after LSM is illustrated in Fig.2.
We have found out that by setting some thresholds and
conditions for these values, the more reliable and accurate
result can be selected from all the results of different
window sizes.
A series of window size 35x35, 39x39 and 43x43 will be
used for the LSM. (The window size can be changed of
course to fit the different conditions like different image
scale, pixel size and image content, etc.. Since the pixel
size is very small in our case, we have started from the
35x35 window size.) If the NCC value of the initial position
found from the previous step is called C1, and the NCC
value after the LSM is called C2, from the matched results
of all possible window sizes, only the one which satisfies
the following conditions will be selected as the final result:
1. Its C2 must be greater than 0.75.
2. Its C2 must be greater than or at least equal to C1.
3. Its C2 must be the greatest among all window sizes.
4. If more than one window size satisfy the above 3
conditions, the one with the largest window size shall be
selected.
When no window size satisfies these conditions or even the
LSM diverges, the system outputs an failure message and
switches to manual selection of target and search window
positions.
6. Test Results
The test block has 4 strips, 26 photos and covers an urban
area with many buildings. The photo scale is 1:5000. For
each tie area approximately 20 TPs are selected. The size
of windows in each pyramid level and the threshold value of
NCC for finding the correspondence are listed in Tab.1. The
thresholds used in the dynamic window LSM are 0.65 for
C1 and 0.75 for C2. For each tie area at most 5, at least 2
triangulation points should be found. Two points are the
minimum to provide sufficient reliability (error detection) for
the subsequent bundle block-adjustment. Tab.2 shows
the number of triangulation points actually matched in
the tie area and the number of tie area sets corresponding
to them. All conjugated areas referred to the same
terrestrial position form a set. A TP in a set of tie areas is
considered as successfully matched only when it is
matched in all tie areas in that set. When it can not be
matched even only in one area, the whole set will be
declared as failed. From Tab.2 we can see that the number
of tie area sets which have at least 2 matched points is 69.
There are 4 sets in which no point could be matched, and 1
Set with only 1 point matched. These five sets are
considered as not successful. There are total 74 sets,
therefore the success rate of matching is 9396. For those
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
