t will be
rogram. This
also on sub
the tie point
task after the
T of relevant
ess relies on
|) importance
|: number of
2 number of
points (good
rement is that
process, the
nber of good
d in the block
1 to checking
ns in each tie
' sufficient. It
servations in
ient. It is not
ons in each of
, to achieve
le images are
lent. The tie
don.
he tie point
are discussed
a Section 3.2.
as
rol have to be
e of difficult
water, forest,
erences etc.).
failed areas,
ure does not
egardless of
is good areas
n the overlap
ere exist no
1 the overlap
nbinations.
hing will not
worst case in
may succeed
^ interactive
To perform the correct action, intelligence is required from the
system. This task is solved interactively in the existing systems.
In the system at FGI, the checking process is also interactive.
The reason for failure is checked and one of the actions
mentioned above is carried out.
2.1.6 Process flow
Knowledge on how the tie point extraction is progressing is
realised in the process flow. Different tasks are evoked using
this knowledge and information gained during the tie point
extraction process.
At the system at FGI the basic process flow is realised at the
moment as follows:
1. Define the proper locations for tie point extraction, see
Section 2.1.1.
2. Extract a large number of tie points in each tie point area,
see Section 2.1.2.
3. Perform block adjustment and select a sufficient number of
points in each tie point area, see Section 2.1.3.
4. Check the quality of the block, see Section 2.1.4.
Process the unsuccessful tie point areas, see Section 2.1.5.
6. Iterate steps 3-5 until the quality is satisfactory. Complete
with final block adjustment.
CA
2.2 About distribution, number and completeness of the tie
point observations
Important factors in the tie point extraction process are
distribution, number and completeness of the tie point
observations. Appropriate values for these factors depend on the
imagery and measurement method used and of course on the
accuracy requirements, but they are not exactly known. They
are briefly discussed below.
2.2.1 Distribution of tie point observations
The concept of the distribution of tie point observations can be
treated on global and local levels. Global distribution means the
distribution of tie point areas on the image. Local distribution
means the distribution of numerous tie point observations in the
tie point area. In the following, global distribution is discussed.
When using conventional aerial imagery and interactive
measurement, it is sufficient to extract tie points in the Gruber
positions (3x3 tie point area distribution). This has been
considered, though not proven, to be sufficient also in the
automatic case, see (Schenk 1995, Tsingas 1992). On the other
hand, the measurement of extra points can easily be carried out
using automatic methods. It is therefore of interest to test if a
more dense distribution of tie point areas will lead to an
increase in the accuracy of the block. A 5x5 distribution on the
images was tested and the results are presented in Section 3.2.2.
2.2.2 Number of tie point observations
The number of observations can be huge in automatic tie point
extraction. The main reasons for this are: 1) it is usually easy to
measure a large number of observations, 2) the quality of the
observations is unknown (matched objects may be poor which
concerns all measurement methods), and better accuracy is
achieved by increasing the number of observations and 3) the
accuracy of some commonly used image matching methods is
poor (FBM).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
a)
Strip 1 Image / Image 2 Image 3
Strip 2 Image 4 Image 5 Image 6
b)
Neighbour- Image combinations
hood
4 «1:2, 4: 5»..«2,3,.5,6»
3 «1,2, dol, 2,52, «l, 4, 55,
«2, 4. 3»,
«2 dm Sad nz,
«3 S 6»
2
inside strip «1/2»,«2,3»|«4,5»,«5, 6»,
between strips | «1, 4», «2, 5», «3, 6»
Figure 1. Splitting a 6-fold tie point area. a) Overlap area: two
strips with 3 images. b) Splitting to 4-, 3- and 2-neighbouring
image combinations.
The question about the number of tie point observations is often
too much simplified: the more observations the better results.
200-300 observations/image seems to be commonly used. There
are evidences that an increasing number of observations does
not necessarily lead to better results. One important reason for
this is that not all observations have any significant influence on
the result. The effect of the number of observations was tested
and the results are presented in section 3.2.1.
2.2.3 Completeness of the tie point observations
As mentioned in Section 2.1.1, to achieve stability in the block,
matches on multiple images are needed. The problem is that
matches especially in 6-fold tie point areas may easily fail. This
is because the overlap area tends to be small and there are often
big radiometric and geometric differences between the
overlapping images, which can not be dealt with using known
image matching techniques, see also 2.1.5.
In some cases tie point areas have to be split. In general, in a n-
n
fold tie point area, there are Sm different image combina-
iz2
tions (for instance, 57 image combinations in a 6-fold area). In
practice, successful matches are usually most likely to be found
between neighbouring images. In Fig 1. splitting a 6-fold tie
point area into 4-, 3- and 2-neighbouring image combinations is
shown. The effect of splitting the tie point observations was
tested and the results are presented in Section 3.2.3.
3. EMPIRICAL INVESTIGATION
3.1 Test arrangements
3.1.1 Subjects studied
The following subjects were studied: 1) selecting a varying
number of points from each tie point area, 2) reducing the
completeness of the observations, 3) using 5x5 tie point area
distribution and 4) using a tie point extraction strategy
combining multiple and pairwise matches. The investigation is
not comprehensive, it is meant to give ideas about the effect of
some factors.
339
SRE
ness
En
TERN EOS Ee
CORP
"s
