project area
entation ele-
x map), the
on, informa-
n of special
Iso assumed
e imagery is
reprocessed.
ned that the
an operator.
e.g., Gülch,
tic methods
verse shapes
able level of
also bear in
rborne GPS
i remark to
:d: it is con-
he block ad-
they can be
1y time.
tions of the
points can be
this purpose
entation pa-
As a rough
be known to
angular ac-
10to scale of
ts to 10 m.
re not so ac-
must be de-
stment with
1995). But
| parameters
urface is not
n quite well,
[oreover, the
ions is inac-
procedure to
, the surface
b
the criteria
he reader is
oint” should
ts a concept
oints, edges,
eulation sys-
the classical
nt cluster in
95; Tsingas,
t these loca-
erlap config-
n. Some ap-
her flat sur-
NS systems.
js may come
Table 4: Minimum and desirable requirements for
solving critical tasks.
essential requirements
tasks minimum | desirable
assumptions
= ext. or. |L-GPS/INS no restrictions
— surface flat no restrictions
selection
- number | limited unlimited
— location | random planned
prediction overlap predictor
matching
— approx. image pyramid | image pyramid
single image all images
— MIM pair wise simultaneous
from processes following aerial triangulation, for ex-
ample DEM generation. Here, the adjusted tie points
(blockpoints) serve as initial seeds. Consequently,
they should be in strategically relevant locations, e.g.,
on breaklines. Working with edges as entities in aerial
triangulation would greatly facilitate the DEM pro-
cess because edges are likely to correspond to break-
lines (see, e.g., Schenk, 1992).
The final step of multiple image matching requires
very good approximations for the matching windows.
This is the purpose of the approximation task. Usu-
ally, a hierarchical approach is preferred where the
selected matching entities are tracked through the im-
age pyramid. At first sight the task appears trivial,
but there are some intricate details that make it a
challenge. It is known from scale space theory that
features selected on one level of the image pyramid
may disappear on higher resolution levels. It is quite
unlikely that features to be matched on the high res-
olution level appear on the coarse level where the hi-
erarchical approach begins. As a consequence, new
features must be extracted on every level, and, more
important, must be matched. It is during this process
that some of the original n-connectivity (object space
feature appears on n images) is lost, thus weakening
the block stability.
5.4 Summary
Some of the more important aspects of automatic
aerial triangulation are summarized in Table 4. The
first column contains essential tasks, while the second
and third columns indicate minimum and desirable
specifications.
To offer the most flexibility it is desirable to have a
system that does not depend on GPS/INS informa-
tion for the initial estimates of the exterior orientation
parameters. A more relaxed assumption is the stan-
dard aerial case where the attitude may reach 5° and
the base elements may vary as much as 10 percent
743
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
from ideal overlap conditions. Presumably, the most
severe restrictions of automatic aerial triangulation
systems are in the surface conditions. As another ex-
ample to demonstrate the importance of knowing (or
computing) the surface, suppose the initial match be-
gins on a 128 x 128 resolution level, pixel size ~ 1.8
mm. For a mountainous area with elevation differ-
ences up to 1/3 of the flying height, the surface un-
certainty amounts to 24 pixels—probably too much
for any matching scheme.
As elaborated in Section 5.1.142, the selection of
blockpoints should be intelligent and unrestricted in
number of points. For predicting matching locations
it is desirable to have a sophisticated “predictor” that
also determines the uncertainty of the estimated con-
jugate locations based on the uncertainties of all pa-
rameters involved, such as exterior orientation and
surface. Sometimes, multiple image matching (MIM)
is approximated by matching all possible pairs of im-
ages. Again, it is desirable to employ a MIM scheme
capable of simultaneously matching all entities. Rig-
orous solutions are described in (Agouris, 1993; Krup-
nik, 1994; Schenk, 1996).
6 CONCLUSIONS
Digital aerial triangulation is here! It comes in two
forms: interactive and automatic. Interactive meth-
ods depend on a human operator who makes critical
decisions and takes over control should the system
fail. Not an absolute necessity, most interactive meth-
ods are built around softcopy workstations. The com-
bination of well established and familiar procedures of
automatic aerial triangulation with image processing
and softcopy workstations results in attractive solu-
tions that successfully compete with traditional meth-
ods. Virtually every softcopy workstation now has a
digital aerial triangulation component.
Automatic aerial triangulation systems strive for re-
ducing the operator involvement. Such systems run
in a batch environment. What began a few years
ago as a rather esoteric research subject is now on
the verge of entering the marketplace. The first gen-
eration systems will offer a high degree of automa-
tion; the only manual task is the identification and
measurement of targeted points (control points, pre-
marked tie points).
At first sight automatic aerial triangulation resembles
traditional methods. For example, the major tasks of
selecting, transferring, and measuring tie points re-
main as well does the block adjustment. However,
their solution is quite different, except for the block
adjustment. Transferring tie points means predict-
ing conjugate matching locations, measuring means
matching. Since selected features appear on more
than two images, multiple image matching is needed.
There are some distinct differences between auto-