device and
ssibility to
iediately on
'Ssing.
dak, 1998)
’D sensor
r PCMCIA
ty up to 54
all standard
)osure, me-
ified Nikon
be applied
wing angle
CCD chip
film.
es were ac-
copter (see
rom terres-
s restricted
> terrestrial
rest side of
objects. In
hort object
copter was
rmitted. In
nagery was
8 mm), de-
two image
mm lens
1 lens
A total number of 49 images was acquired during the
project. 19 images were taken from the helicopter with the
28 mm lens. These images have an image scale of about
1 : 2850 and the average object distance for theses images
is about 80 m. A total of 30 images was acquired from ter-
restrial viewpoints using the 18 mm lens. These images
have a mean object distance of about 25 m, which results
in an image scale of about 1 : 1400. An overview of the
camera configuration is given in Figure 4.
—L- "aerial" images
— "terrestrial" images
Figure 4: Camera configuration
3. OBJECT ORIENTED MEASUREMENT WITH
DIPAD
The reconstruction of man-made objects is a non-trivial
task. They are often complex, irregular, appear different
according to their function or context, etc. In a typical
non-controlled environment like outdoor scenes, their ex-
traction from imagery is difficult due to occlusions (from
other objects or due to perspective projection), illumina-
tion effects (shadows or weak contrast), radiometric inter-
ferences or varying background.
DIPAD aims on the automated and object oriented genera-
tion of as-built CAD models. It combines a CCD-sensor
based image acquisition with a semi-automatic processing
of the image data in a CAAD controlled environment. The
main features of the system are:
« possibility of self-diagnosis (quality control),
potential for high accuracy and reliability (redundant
sensor data),
e flexibility with respect to the three-dimensional
reconstruction of buildings or parts of buildings, and
e performance of object oriented measurements.
The problem of object recognition and measurement is
solved in a way that the image interpretation task is done
by the user (architect, art historian, etc.) and the recon-
struction and measurement of the precise geometry from
multiple images is performed automatically by the com-
puter (HICOM principle). A human operator makes easily
use of his knowledge about the real world (mostly subcon-
sciously) while looking at an image and can easily filter
the necessary information from the images for his task
and/or complete the missing information in his idea. Fol-
lowing a combined top-down and bottom-up strategy a
coarse given CAD model will be iteratively refined until
the desired degree of detail is achieved.
3.1. Object Models
Object models can be treated as abstractions of real world
objects. These are necessary in order to process objects of
the complex and extensive reality in a computer environ-
ment. Each attempt to represent reality is already an ab-
straction. The most important role played in the definition
of models is the proper balance between correctness and
tractability, i.e., the results given by the model must be ad-
equate both in terms of the solution attained and the cost
to attain the solution.
There are several ways to describe an object in a CAD en-
vironment. In general, 3D models can be divided into
three different classes of models: wireframe models, basi-
cally defined through vertices and their connecting edges,
surface models, describing objects as an ordered set of
surfaces, and volumetric models, describing objects by
volumes. The class of volumetric models as the most in-
teresting one comprises more sub-classes, such as para-
metric models, sweep representation schemes, cell
decomposition schemes, boundary representations, con-
structive solid geometry, hybrid models and others. Each
of these classes has its specific advantages and disadvan-
tages for different tasks. But there is no class which is op-
timal for all tasks.
The formal data structure in DIPAD (see Fig. 5) consists
of two data sets, the photogrammetric data, which con-
tains all the information about cameras, images and sta-
tions, and the object data. The object data consists of three
related data structures: the geometric data, the topologic
data and the thematic data of the object. The topologic
part of the object model consist of six classes, which rep-
resent the hierarchical structure of the object. The ele-
ments of a hierarchical class consists of the elements of
the next lower hierarchical class. The six classes for the
topologic data are vertices, edges, areas, volumes, con-
structive units and objects. The three classes of geometric
primitives contain the geometric description of the corre-
sponding topologic element. These classes are points,
lines and surfaces.
Beside the topologic and the geometric classes there are
also five classes of thematic attributes, which correspond
to the topologic primitives. These attributes contain infor-
mations which are not of topologic or geometric nature
273
