is with the
ers and the
imeras ap-
d through
controlled
el into the
ojection of
ith the ap-
et a visual
modelling
he initial
3.3. Automatic Measurement
The automatic feature measurement employed in DIPAD
is guided by the topologic object description created by
the object modeller or by known a priori CAD informa-
tion about the object and calculates simultaneously the
best match of the elements described by the object model
with the image data of multiple images.
Due to the guidance of the measurement by the topologic
object model, only relevant features (as defined by the us-
er) are extracted and redundant or useless information is
reduced to a minimum. This strategy follows the native
theory of human perception (Gibson, 1950), where infor-
mation is defined by regularity in contrast to coincidence
and not by contents and meaning. This means that a hu-
man being can identify a signal much easier the more re-
dundant the signal is. The same counts for computer
algorithms. In addition the use of a priori knowledge
makes explicit assumptions, that allows the checking of
whether or not these assumptions are fulfilled in the imag-
es. The three-dimensional position of the object is derived
by a simultaneous multi-frame feature measurement,
where the object model is reconstructed and used to trian-
gulate the object points from corresponding image points.
CAD model
1. digital images a
"1. 3D object description in CAD T
p dH [n] processing steps
Figure 8: Processing steps in DIPAD (MARE - auto-
matic measurement of architectural elements)
The whole process of object reconstruction can be divided
into three processing steps, which run consecutively (see
Fig. 8). In the first processing step the simultaneous deter-
mination of the object geometry in image space based on
the radiometric information in the digital images and the
topologic information about the object feature is per-
formed. The routine uses straight lines as the basic enti-
ties, by first locating the edges of the features to be
measured and then deriving the vertices as intersections of
appropriate lines. In the second processing step the pre-
cise geometry of the architectural features in object space
is determined. Therefore the image coordinates of the first
processing step are used to estimate the object coordinates
of the architectural feature. In the case of unknown or ap-
proximate camera parameters these data are estimated as
additional unknowns during the bundle adjustment. The
geometric improved features are reprojected into the im-
ages in order to restart the first processing step. This itera-
tive procedure continues until the position of the feature in
object space after each loop is stable. The third processing
step enables the user to increase the degree of detail for
the topologic object model. Here more object details can
be added to the measurement routine. A more detailed de-
scription of the processing steps is given in (Streilein,
1994) and (Streilein, 1996).
An example for the performance of the first processing
step on a window feature of the monastery is given in Fig-
ure 9. The feature with its approximate geometry is pro-
jected with the approximate camera parameters into the
image and used as starting value for the automatic meas-
urement. The linear feature boundaries are extracted and
straight lines are fitted to the linear feature boundaries. Fi-
nally the image coordinates of the vertices are calculated
by line intersections.
Figure 9: Example for the performance of the first
processing step on a window feature
(a) approximate feature position
(b) extracted linear feature boundaries
(c) straight line fitting to linear feature boundaries
(d) vertex computation by straight line intersection
4. RESULTING MODEL
The final result of the processing of the digital images
with DIPAD is a topologic and geometric object descrip-
tion of the monastery in the CAD environment. Figure 10
shows the resulting model from different view points and
in different representations. The model exists of about
1'800 object points and 1'200 geometric entities.
The 3D geometry of the object was derived by a free net-
work bundle adjustment with self-calibration. The system-
atic errors of still video cameras employing off-the-shelf
lenses with large distortions from the ideal perspective
transformation are accounted by extending the colinearity
equations with functions of additional parameters. Many
additional parameter sets have been developed to meet
various requirements, in close-range CCD-sensor based
systems a set of ten additional parameters has proven to be
effective (Beyer, 1992). These parameters are three chang-
es for the elements of the interior orientation, a scale fac-
tor in x direction. a shear factor, the first three parameters
of radial symmetric lens distortion and the first two pa-
rameters of lens decentering distortion.
The processing of the 3D data was performed in two
steps. In a first step the parameters of exterior and interior
275
