PROCESSING OF PROTRUDING OBJECTS IN DIGITAL IMAGES OF URBAN AREAS
Borislav D. Marinov
University of Architecture, Civil Engineering and Geodesy, Faculty of GOEDESY
BULGARIA
Commission III, Working Group 3
KEY WORDS: Interpretation, Urban Image, Scene Modelling, Computer Vision.
ABSTRACT
For purposes of automatic processing of digital stereo images of urban areas and architectural objects a skeleton space model is
introduced. The topological and height properties are used for classification of such objects. The image areas are separated in several
main classes: visible terrain, building roofs and building walls. A formal approach for creation of such model is introduced based on
picture language description procedure. The diversity of relative conjunction of areas in every photo image is modelling by applying
the sequential grammar rules based on PLEX-GRAMMAR. The skeleton model of photo images is suitable for purposes of terrain
building separation process and for finding the correspondence between visible areas in two stereo images. The possible areas of
application of such model are interpretation and verification of skeleton model of urban areas, the rejection of building influence
over the produced DEM and appropriate selection of image parts for digital image mosaic creation from large scale images in
processes of photo transformation and land cover classification.
1. INTRODUCTION
Digital processing of large scale images of urban regions
reaches severe problems due to the high density of buildings
with the different heights from the surrounding terrain. For such
objects there are a great displacement of their photo-images,
hiding parts of the surrounding terrain, mutual hiding of close
staying buildings. Similar problems exist in close-range images
of architectural objects with a lot of protruding elements on the
facades. As a result a problem arises in the terrain interpolation
or the orthophoto image generation for such parts of the digital
images where discontinuity of the terrain function appears.
Automatic creation of normalised classified images from such
areas is not enough reliable and complete due to the
unclassified areas in the hidden parts of the terrain. There are
suggested different schemes for classification and processing of
such images, taking into account geometrical, shape and height
properties of buildings (Weidner, Fórstner, 1995). For object
with simple form a formalised procedures for scene analysis
and image understanding in single image are used, based on
skeleton model of objects (Waltz, 1975).
One possible solution of problems with processing of urban
region images is a detailed mapping from stereo images and
building contours coding under the operator's control. The
building's boundaries could be excluded by defining a special
type of lines for such objects - like border lines that allow to
exclude the undesirable areas from the process of DEM
generation). Such procedure seems to be very tedious and time
consuming for high density built urban areas. Such direct
approach could not be enough appropriate in cases of hiding
some parts of the buildings or in very unusual relative
disposition of objects because very complicated image
understanding procedure have to be applied.
2. IMAGE DESCRIPTION
Another solution is to introduce a simplified model of stereo
images. The description is based on the topological properties
497
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
of areas in the stereo images. A picture language grammar
technique could be applied to formalised the process of image
description and to cover different possibilities of image
configurations. Such technique is very largely used for
description of two dimensional or three dimensional images
(Fu, 1982). Similar approach was suggested for description of
planar objects in single photos or contour maps (Marinov,
1988). It formalises the process of generating hierarchical
description of contour images. For purposes of stereo image
processing such approach have to' be enlarged to involve
additional information. The information includes the parallax of
contour lines, the position of the objects relatively to centre of
the photo and the topological properties of areas in both
images. In that situation only the walls oriented to the centre of
the photo could generate visible areas in the image. The main
properties of the image description have to take into account
the relation between the areas in the image and connection
between the areas and the contours. Primary elements that are
used for description are arcs (the segments of lines between two
points of connection). From a topological point of view nodes
are only points of connection of more than two arcs, taking into
account the type of arcs (wall edges, foundation border and so
on) the segmentation of arcs could increase. The main
topological properties that have to be taken into account are the
relations of the types: adjacent areas, clusters of areas, included
clusters of areas, hanging areas (with only one point of
connection. The specific properties of artificial objects includes
areas of following types: roofs, walls, terrain and hidden parts
of them.
For purposes of contour description only the tie points and the
arcs could be used. In such situation it would be difficult to
generate relation between areas. The usage only of the
topological properties of areas and their relation would increase
the difficulties in contour generation and arcs processing.
The form of description proposed here is based on defining the
different types of arcs elements and topological description of
relations between them. The primary topological elements are
arcs separating adjacent areas. There are involved four types of
