CAPES
e use of models for
of 3D —geometry is
s e.g. that roads are
? about the appear-
y occluded objects.
ly by semantic nets
he expectations of
s of forests exhibit
Ience, to improve
system uses prior
econstruction. To
object classes like
erpretation is re-
regions. Consecu-
ol the 3D —recon-
se for new model-
nulated explicitly.
1 by formal logic,
|, 1993), semantic
s, rule based sys-
) and neural nets
ural relations se-
ed by the work of
ire employed for
the system archi-
d methods for ex-
nsecutive chapter
'retation. Chapter
cription is used to
er concludes with
Symbolic Scene
Description
Landscape—1
Sensor Data
Fig. 1: Architecture of knowledge based modelling system AIDA
2. SYSTEM OVERVIEW
Figure 1 shows the architecture of the knowledge based mod-
eler AIDA. The goal of the modeler is a realistic reconstruc-
tion of the observed scene. Input to the modelling are over-
lapping aerial images and prior knowledge about the objects
present in the scene. Modelling consists of three main mod-
ules.
Image processing: The overlapping aerial images are recti-
fied in a way that the epipolar line coincides with the image
scanline to ease search for homologous points. Consecutively
a height map is computed from the stereoscopic image pair.
Further line shaped features and regions are segmented in
the image.
Symbolic processing: Interpretation uses knowledge about
the expected objects to group the features and assign a scene
specific semantic to them resulting in a symbolic scene de-
scription.
Vector processing: From object semantic geometric
constraints are derived to restrict the free parameters of sur-
face reconstruction. The objects are approximated by a sur-
face mesh with overlaid photo texture.
3. KNOWLEDGE BASE
3.1 Types of Knowledge
The a priori knowledge for 3D reconstruction of landscapes
from aerial images includes knowledge about
e Objects,
e context and task,
e sensors, and
e strategies.
Objects possess attributes and relations to other objects. As
attributes geometry (e.g. shape, size, etc.), material (e.g. con-
crete, sand, etc.), and function can be distinguished.
869
INPUT
Knowledge Base Symbolic Processing:
Landscape INTERPRETATION
Forest Road
Geometric Constraints
Vector Processing:
RECONSTRUCTION
Features
Image Processing:
SEGMENTATION
HEIGHT ESTIMATION
Forest—1 Road—1
3D —Surface Model
Objects appear only in special contexts, i.e. forest edges in the
context of forests. The task specializes the modelling de-
mands. Both, context and task, reduce the problem domain.
Sensors transform objects into another, here pictorial repre-
sentation, using geometric and radiometric transform char-
acteristics. Image processing operators can be regarded as
sensors that transform images to images. Their representa-
tion is not within the scope of this paper. For a representation
of image processing knowledge the reader is referred to the
system CONNY (Liedtke, 1992).
Strategies state how and in which sequence scene analysis has
to proceed. E.g. eminent objects have to be searched for first.
3.2 Knowledge Representation
3.2.1 Objects: Object representation employs frames which
contain a collection of attributes, relations, and methods (fig.
2). The relation slot establishes the connection to other ob-
jects. The object properties are stored as attribute values.
Further the object has methods, i.e. functions, at its disposal
to compute the attribute values. There may also be a method
available to segment the object in the image data.
Main Road
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Relations:
is—a: Road
part—of—inverse: Road Segment
Attributes:
Width[m]: 10...20
Material: Asphalt
Methods:
Segmentation: RoadExtractionFunction
Fig. 2: Example for a frame
