KNOWLEDGE BASED MODELLING OF LANDSCAPES
Ralf Tönjes
Institut für Theoretische Nachrichtentechnik und Informationsverarbeitung,
Division "Automatic Image Interpretation", Chair Prof. Dr.—Ing. Liedtke,
Universität Hannover, Appelstr. 9A, D—30167 Hannover,
WWW: http://www.tnt.uni—hannover.de/ ~ toenjes.html
E—mail: toenjes@tnt.uni—hannover.de
KEY WORDS: Knowledge Base, Interpretation, Modeling, DEM/DTM, Stereoscopic, Visualization
ABSTRACT:
A knowledge based approach for automatic generation of 3D —landscape models from aerial images is presented. The use of models for
visualization tasks results in two requirements: efficient representation and high realism. Efficient representation of 3D —geometry is
achieved by polygon meshes. Realism requires that the models meet the expectations of a human observer, who knows e.g. that roads are
planar and forest edges possess a height step. The presented knowledge based modeler AIDA employs prior knowledge about the appear-
ance of the objects in the scene to derive object specific constraints for surface reconstruction and to complete partially occluded objects.
This requires an image interpretation to assign a semantic to the scene objects. The knowledge is represented explicitly by semantic nets
and rules.
1. INTRODUCTION
For visualization of synthetic scenes 3D—models are
required from which new simulated views can be computed.
Applications such as flight and driving simulators, movie and
TV production have a high demand for realistic models. Es-
pecially Landscape visualization is becoming an important
tool for earth scientists, environmental researchers and civil
engineers. Quantity, precision and the kind of models ask for
methods that automate the model generation.
The common approach for digital terrain modelling uses ste-
reo matching techniques to recover the height information
from aerial images [Ackermann, 1991]. For efficient visual-
ization the height map is subsampled and approximated by a
polygon mesh in space. The geometric and photometric fine
structure is modelled by projecting the aerial images onto the
polygon surfaces.
However, the reconstruction of a 3D— model from its 2D—
projections is an inverse and underconstrained problem,
which causes model errors:
e The model is incomplete due to occlusions. This applies
often to edges of forests and roads passing through fo-
rests.
e The sensor resolution limits the level of detail for recon-
struction. Especially missing height steps in connection
with shadows appear erroneous.
e The mesh approximation of the model geometry does
not correspond with the object boundaries resulting in
faulty breaklines.
868
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
This yields often models that do not meet the expectations of
a human observer, who knows that the edges of forests exhibit
a height step and roads run continuously. Hence, to improve
the realism of the model the presented system uses prior
knowledge about the landscape for 3D— reconstruction. To
exploit the prototype knowledge about the object classes like
roads, forests and grassland an image interpretation is re-
quired that assigns a meaning to the image regions. Consecu-
tively the object semantic is used to control the 3D—recon-
struction.
To ease the adaptation of the knowledge base for new model-
ling tasks, the knowledge base has to be formulated explicitly.
In general knowledge can be represented by formal logic,
fuzzy logic, frames (Clement, 1993; Foresti , 1993), semantic
nets (Niemann, 1990), production systems, rule based sys-
tems (Matsuyama, 1990; Mc Keown, 1985) and neural nets
(Shapiro, 1992). For description of structural relations se-
mantic nets are suited. Hence here, inspired by the work of
Niemann et. al. [1990], semantic nets are employed for
knowledge representation
The following chapter gives an overview of the system archi-
tecture. The third chapter presents the used methods for ex-
plicit knowledge representation. In the consecutive chapter
the knowledge base is used for image interpretation. Chapter
five describes how the symbolic scene description is used to
improve the object reconstruction. The paper concludes with
a presentation of the results.
Fig
Fig
elei
tior
lapi
pre
ule:
Im:
fiec
scal
a he
Fur
the
Syn
the
spe
scri
Vec
con
faci
fac
atti
cre
