1ethods and
ching
Lies
levels
ional
iption
npares sym-
the similar-
iptions may
. They can
itic nets. In
matching is
properties.
a similarity
es. Feature-
nore robust
ngulation is
neously. We
n more than
1) aerial tri-
re consider-
| interactive
tion. Gen-
for manual
The differ-
efined; both
ss where the
ge, and the
allowing the
ver, a much
-the subject
e way aerial
lotters. Ad-
laying more
r identifica-
1easurement
iage match-
a controlled
all processes
'henever the
matic" that
ate that the
point measurement is performed "automatically" by
employing a matching algorithm. In the next sec-
tion the term is used in a much broader sense: not
only are the points measured automatically, but also
their selection, their transfer, and the determination
of suitable approximations.
A major advantage of interactive aerial triangulation
is of a practical nature: users may follow familiar pro-
cedures thus reducing the risk of making costly mis-
takes with new technologies.
Various interactive aerial triangulation systems are in
practical use and results about performance and ex-
perience haven been reported (e.g., Haumann, 1995;
Beckschäfer, 1995).
5 AUTOMATIC AERIAL
TRIANGULATION
Automatic aerial triangulation systems are on the
verge of entering the marketplace. Several systems
have been described (see, e.g. Ackermann, 1995;
Krzystek et al., 1995; Schenk, 1995). Tsingas (1995)
and Fritsch (1994, 1995) report about experimental
results. Some systems evolved from successful solu-
tions of automating the relative orientation (see, e.g.,
Tang et al., 1994; Mayr, 1995).
With automatic aerial triangulation we mean meth-
ods that attempt to solve the task as a batch pro-
cess, with little or no help from a human operator,
except the measurement of control points. In order
to achieve this ambitious goal it is imperative to fully
exploit the potential of digital photogrammtry, im-
age processing, and computer vision. This, in turn,
may suggest taking a fresh look at the problem rather
than mimicking existing procedures that are optimal
for traditional aerial triangulation but perhaps not
ideally suited for an automatic approach. We pursue
this view here and derive essential tasks from the ob-
jectives of automatic aerial triangulation. These es-
sential tasks must be addressed by every aerial trian-
gulation system in one way or another—they are sort
of invariant, independent of the method. Incidentally,
the solution of essential tasks determines the level of
comfort and performance of automatic aerial triangu-
lation systems. Thus, they may serve as evaluation
criteria.
5.1 Essential Tasks
The essential tasks of automatic aerial triangulation
are derived in a backward fashion, starting from the
objectives which include the determination of the ex-
terior orientation parameters and a partial recon-
struction of the object space. We request that the
orientation parameters are as accurate and reliable
as in analytical aerial triangulation. The adjusted
blockpoints form a minimal reconstruction of the ob-
ject space. With blockpoints, or tie points, we refer
739
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
to the matching entities. They may comprise points,
point features, or line features. More advanced Sys-
tems would also include regions, ranging from small
surface patches with their topography and radiometry
to the entire surface of the project area. Obviously,
this latter concept would combine aerial triangula-
tion, DEM and orthophoto generation in one process.
5.1.1 Number of Blockpoints: With analyti-
cal methods accurate orientation parameters are ob-
tained by measuring relatively few blockpoints, say, 9
to 15 points per photograph, as precisely as possible.
For economic reasons the challenge is to work with a
minimum number of points that still assure reliable
and accurate results. Consequently, the points must
be carefully chosen, transferred, and measured.
In automatic aerial triangulation the situation is quite
different. First, it really does not increase computing
cost significantly if hundreds of points per image are
matched. But more important, human operators are
far superior in selecting blockpoints than machines.
Âs a result, we are much better off in using many,
but less carefully chosen blockpoints. The empha-
sis shifts from a few points to masses of points. Isn't
this neglecting the accuracy aspect? Suppose we have
25 times more blockpoints in automatic aerial trian-
gulation than in the standard case. Since all points
contribute to the determination of the exterior orien-
tation parameters, their accuracy will be roughly five
times better. This is the same as to say that the accu-
racy of the orientation parameters remains the same
if many but less accurate points are used. The com-
pelling conclusion is that the accuracy of an individual
point is much less important than in traditional aerial
triangulation. Consider a pixel size of 30 um and a
matching accuracy of 1/3 of a pixel. Even though this
measuring accuracy is not outstanding at all we still
obtain more accurate orientation parameters. More-
over, the reliability increases.
The claim of higher accuracy and reliability of the
exterior orientation parameters is confirmed by ex-
perimental results (see, e.g., Fritsch, 1995; Tsingas,
1995; Ackermann, 1995).
9.1.2 Location of Blockpoints: Even though the
selection of blockpoints is less critical than in tradi-
tional aerial triangulation, their location should not
be arbitrarily chosen. This is particularly true if aerial
triangulation is viewed as a preprocess to other pho-
togrammetric procedures, such as DEM generation
and map compilation. For example, the automatic
generation of DEMs would greatly benefit if the block-
points were selected at interesting locations, such as
along breaklines. In any case, the selection should
satisfy the following criteria:
multiple overlap object points that appear on as
many images as possible increase the stability of
the block adjustment.
