ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001
416
description of the scene (objects, geometry, colours), which
is very appropriate for the rendering system.
Accuracy. The accuracy of the re-constructed 3D model
plays an important role in the entire process. All the objects
that are approachable by the user within few meters
distance have to be reconstructed with an accuracy that
corresponds to the accuracy of real-time extracted lines
from the video images. The user might get as close to an
object as one centimetre may be of significance. Indeed,
aiming at centimetre accuracy of the objects is a very
expensive and practically an unrealistic requirement.
However, certain parts, elements or section of real objects
have to ensure accuracy ranging from few centimetres to a
decimetre. Doors and windows at street level, balconies at
lower floors, statues, paths, tiles (see Figure 9), etc. are
examples of such elements. The re-construction methods
applied to achieve such precision are given in section 4.
Siili
i .;. 4>
m:
Figure 2: Geometry details represented by line features (the
façade in the middle)
Analysis of the requirements reveals the complexity of the issue
leading even to a contradiction, i.e. high level of details and
maintenance of 3D topology. Apparently, the construction and
maintenance of complete 3D topological model (including
windows, doors, etc.) is time and effort expensive and therefore
inappropriate. One approach (utilised in our project) is a clear
discrimination between the types of data needed for the two
subsystems. The line features (expected from the positioning
subsystem) are kept as straight lines (see Figure 2). The
outlines of the 3D objects are considered a separate data set
and organised in a 3D topological data structure (see Figure 3).
Thus the amount of data to be topologically maintained can be
significantly reduced that contributes to the simplification of
many spatial operations and thus to the speeding up the data
retrieval. Furthermore, the two data sets are linked by explicit
reference between line features and façades. The number of
lines is expected to be rather large (for one façade, it may rise to
300-400) and therefore the 3D reconstruction process takes
care relationships “a line belong to a face” to be created and
explicitly stored in the database.
3D DATA STRUCTURING
The research in 3D data structuring attracts a lot of attention in
the last several years. The focus is basically on an extended
conceptual model capable of integrating geometric (position,
shape and size) and thematic characteristics of objects, and
mutual spatial relationships. One stream of investigations
emphasises on formalism (structure, ordering and operators) to
construct a geometric object regardless of the dimension (see
[10]). Such models aim at the complete representation of all the
topological relationships among the objects from different
dimensions. The models can be referred to as an implicit
representation of objects, i.e. the relationships are stored and
the description of the objects can be derived out of them. The
disadvantage is the size of the database that grows
tremendously with the complexity of the model. Many reported
3D models give priorities to topological models that maintain
objects (i.e. an explicit description of objects). More details on
data structures of this group can be found in [3], [5], [9], [15].
The major problem of such 3D models is that a few of them are
tested on large data sets under real-time requirements.
An intensive work on clarifying guidelines for developing GIS
and database software (maintaining 2D, 3D topology) is carried
out as well. OpenGIS specifications are one of the commonly
accepted standards (see [6]). The approaches proposed there,
however, are based on separate maintenance of geometry and
topology objects, which in practice leads to large duplications.
Furthermore, the most of the topological models proposed for
implementation consider mostly the 2D world.
Figure 3: Outlines of 3D objects
Bearing in mind the requirements to the model delineated in the
previous section and utilising recent achievements in 3D GIS
research, we propose a 3D topological model extended to
provide data to augmented reality application. The proposed 3D
model is a typical boundary model with explicit description of
objects. To represent its geometric properties related shape,
size and position, a spatial object can be associated with four
abstractions namely point, iinestring, surface and polyhedron
(see Figure 4). The notations of the four abstract objects
correspond to the ones accepted in the OpenGIS specifications.
A point is an object that does not have shape or size but
position. A linestring is a type of an object that has length and
position. A surface is an abstraction of object that has position
and area. A polyhedron has a position and a volume. These
objects are built of smaller, simpler elements, called node and
face. Nodes describe spatial objects that can be represented as
linestrings (e.g. pipe lines) and points (e.g. trees, lampposts).
Nodes are constructive elements of faces as well. The order of
the nodes in the face is maintained as wheel. The orientation of
the faces is anticlockwise looking at the objects (e.g. buildings)
from outside. Faces are to be used for the reconstruction of
objects that are associated with surfaces (e.g. streets, parking
lots) and polyhedrons (e.g. buildings).
Figure 4: Examples of spatial objects to be supported
Every 3D object can be represented by its thematic and physical
characteristics as well. For example the spatial object building
has a year of building, owner and usage that is referred to as
thematic characteristics. Physical properties are related to
surface reflectance that determines the colour and texture of the
objects. Although the model is potentially capable of maintaining