of the adaptive
atching can be
od is also sim-
an example to
oe checked for
agnitude of its
because there
ve images, i.e.
iar adjustment
parameters of
> of the points
:d to judge the
priori variance
neters without
ntation param-
:oplanar equa-
can be applied
best matching,
orgent images.
tation method
1995].
on the feature
ions must be
| image is usu-
mation volume
d. The Gauss
ng will be only
hich should be
moothing etc.)
tted and new
mid levels until
je features are
the structural
r/Gülch, 1987]
point accuracy
xtraction. The
ir coordinates,
In order to get
ons, there are
10:
line
line
oints is also a
ince the image
oss correlation
en developed
directions.
Image edges are firstly extracted by means of a one-
dimensional operator, which is developed for the fast and
efficient extraction of edges. The lines, which connectivity
must be unique, are extracted from the edges with the
mathematical morphological transformation [Wang, 1994].
The attributes of a line could be the begin and end point,
average gray value, line length, line strength, and line cur-
vature etc. A line have also defined relations with other
lines and regions.
The regions are extracted with the methods of image seg-
mentation. A boundary lines based region growing method
has been developed for the region extraction, which bene-
fits from the advantages of both contour-oriented methods
and region growing methods [Wang, 1992 and 1994]. A
region can be described with reference coordinates, region
size, boundary size, gray value, variance, region form etc.
The region form can be presented with the moment coeffi-
cients or the Fourier Transformation coefficients.
There are five kinds of relations considered in the work.
They are point to line, point to region, line to line, line to
region and region to region relation. the geometric rela-
tions (e.g. angle of two crossed lines) are treated similarly
as the feature primitives, and the topological relations (e.g.
T-crossing) are used as the constraints.
The work flow of the structural matching can be demon-
strated with the Figure 4. With the success of the structural
naeh the corresponding image points of digital stereo
sub-struct. pro ability
"pyramid generation | sub-structures ordering f
image preprocessing
geometry-constrained :
adaptive tree search €
pyramid bac matching |
new points densificationf
attributes of Ie i
| Fig. 4: flow chart of structural matching of two images
images can be recognized fully automatically without
knowing any a priori information such as image overlap
and image orientation parameters.
3. APPLICATIONS
The structural matching methods can be applied in the
cases, where a correspondence between two data
descriptions should be found, e.g. two- or three-dimen-
921
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
sional object recognition between images and maps or
models, object reconstruction etc. In photogrammetry the
problems such as the automatic relative orientation, the
automatic data acquisition for DTM, the automatic aerotri-
angulation, the automatic recognition of control points and
other describable objects can be completely solved by the
structural matching methods.
Since a few years a program system for the photogram-
metric automation has been developed under Microsoft
Windows and is transferred to a Silicon Graphics machine
at our institute. Its aim is to solve the photogrammetric
tasks such as automatic orientation, triangulation and sur-
face reconstruction with highest automation grade. The
aim is reached by using the structural matching method.
Following are two examples for the application of the struc-
tural matching in automatic orientation and triangulation in
photogrammetry with the developed program system. The
data acquisition for DTM and surface reconstruction can
also be fully automated with the structural matching, even
though from the non-metric images or line scanner
images. Due to the limited page number the corresponding
examples will be given in other papers [e.g. Wang, 1994].
3.1 Fully automatic relative orientation
Figure 5(a) displays a stereo image pair in close range
photogrammetry. It is a convergent pair with up-down con-
figuration of the photographic centers. Without to know
XN
{
Emu
*
Fano =
Amn soll
TTY Nem ~~ = A
um A em m
Ameer
Fig. 5(c): recognized lines by structural matching
any other information except digital images the corre-
sponding image points can be recognized by the structural
matching and the parameters of the coplanar condition
can be computed with the developed linear method. Figure
5(b) shows the extracted feature lines for the structural
matching. Figure 5(c) shows the recognized feature lines
on the images by the structural matching. The five conven-
c t c d
