the proven concept of matching two image patches at
the same time. In that case, all possible pairs are
matched, perhaps even forth and back as a control
(see, e.g. Fritsch, 1995). This approach has several
disadvantages. Fore one, the pairs are not indepen-
dent from each other. Moreover, it may happen that
the sequential matching procedure comes to an early
end, leaving some alternatives unexplored.
The information content in every image patch can
only be fully exploited if all patches are simulta-
neously. Agouris presents a rigorous solution in
(Agouris, 1992) that eliminates the problems of se-
quential approaches. A different approach is proposed
in (Krupnik, 1994; Schenk et al, 1996). Here, the
matching is performed in object space. The exte-
rior orientation, the topography of the surface patch
and its gray levels are the parameters to be deter-
mined. Matching in object space has been proposed
by other researchers (see, e.g., Ebner et al., Wrobel,
1987; Heipke, 1990). Fórstner (1995), elaborates on
the differences between these approaches. All these
methods use gray levels as matching entities and the
matching method is LSM. Tsingas solves the multiple
image matching problem in a different fashion (Tsin-
gas, 1994). Instead of gray levels interest points are
used as matching entities and the matching method
is based on graphs.
5.3 Work Flow
We identified several problems automatic aerial tri-
angulation systems must solve in order to deserve the
predicate “automatic.” There is considerable flexibil-
ity in the solutions, resulting in different levels of com-
fort and performance.
iles
selection of tie points
initial assumptions
approximations for
MIM windows
!
multiple image matching
determine approximations
for EXOR and DEM
1 Y
| final block adjustment
determine footprints
(Girl dns
Figure 5: Workflow of automatic aerial triangulation
system.
The workflow in Fig. 5 is a schematic diagram that
depicts the major steps an automatic aerial triangula-
tion system must take. A real system may omit some
of the steps or follow a different sequence. At any rate,
every system begins with some initial assumptions.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
This may involve the topography of the project area
(e.g., flat), the availability of exterior orientation ele-
ments (e.g., GPS/INS system, flight index map), the
block geometry (e.g., overlap configuration, informa-
tion about cross-flights, holes, delineation of special
areas, such as lakes). Furthermore, it is also assumed
that the interior orientation is known, the imagery is
controlled and perhaps radiometrically preprocessed.
For the immediate future it must be assumed that the
control points will be measured by a human operator.
Despite encouraging experiments (see, e.g., Gülch,
1995), there is little hope that automatic methods
would soon cope with recognizing the diverse shapes
of targets in noisy images to an acceptable level of
confidence and reliability. One should also bear in
mind that with the increasing use of airborne GPS
fewer control points are required. A final remark to
when the control points must be measured: it is con-
ceivable to perform all steps, including the block ad-
justment, without control points. Thus, they can be
measured and added to the process at any time.
As argued in 5.2.1 fairly accurate locations of the
footprints must be known so that tie points can be
selected in highly overlapping areas. For this purpose
good approximations of the exterior orientation pa-
rameters and of the surface are essential. As a rough
estimate the exterior orientation should be known to
1-2 mm in image scale. This implies an angular ac-
curacy of about half a degree. For a photo scale of
1:10,000 the positional accuracy amounts to 10 m.
Quite often the orientation parameters are not so ac-
curately known at the outset and they must be de-
termined, for example, by a block adjustment with
coarse measurements (see, e.g., Schenk, 1995). But
even the most precise exterior orientation parameters
do not render accurate footprints if the surface is not
known. In fact, the surface must be known quite well,
otherwise the footprints will be wrong. Moreover, the
prediction of conjugate matching locations is inac-
curate, perhaps causing the matching procedure to
fail (outside pull-in range). In conclusion, the surface
should be known to approximately 3 mm.
The selection of tie points should follow the criteria
sketched in Section 5.1.2. Incidentally, the reader is
reminded that the well-known term “tie point” should
not be taken too literally here: it reflects a concept
that includes features, such as interest points, edges,
and regions. Most automatic aerial triangulation sys-
tems work with a regular pattern, say, the classical
9 point locations, and determine a point cluster in
these locations (see, e.g., Ackermann, 1995; Tsingas,
1992). However, it must be stressed that these loca-
tions must be determined from precise overlap config-
urations, surface and exterior orientation. Some ap-
proaches take a shortcut and assume rather flat sur-
faces and orientation data from GPS/INS systems.
Another important request for tie points may come
742
soh
|
= tel
fror
am]
(blc
the
on |
tria
cess
line
The
very
Thi
ally.
sele
age
but
chal
feat
may
unli
olut
erar
feat
imp
that
feat
the
5.4
Som
aeri
first
and
spec
To «
syst
tion
para
darc
the
