INTERPRETING SINGLE IMAGES
OF POLYHEDRA
Carola Braun
Institute of Photogrammetry
University of Bonn, FRG
abstract
The paper presents a concept of a rule-based system for
deriving the 3D structure of polyhedral objects in perspective
projections using known geometrical constraints. The system
may supply a complete reconstruction of objects, especially of
buildings in aerial images, by matching the results of
individual image interpretation in space.
As buildings can be simplified to polyhedra, the procedures
first are developed for polyhedral objects. Many real objects
have groups of parallel and orthogonal lines and rectangular
corners. The presented system uses this information for
reconstructing polyhedral objects with the aim of a
geometrically correct shape description. The geometric
reasoning process, implemented in the system, is able to
automatically determine the structure of the object step by step,
asking the operator to provide information about geometrical
constraints if necessary. Examples illustrate the applicability of
the approach.
keywords
rule-based system, inverse perspectivity, perceptual grouping,
geometric reasoning, image analysis
1 Introduction
The concept of a rule-based system is presented for deriving
the geometrical shape of polyhedral objects in perspective
projection using known geometrical constraints. The system is
intended to be used for a complete reconstruction of buildings
in aerial images, matching the results of individual image
interpretations in space.
The system is part of a long term project for the development
of procedures for the recognition and reconstruction of
buildings.
As man-made objects like buildings may reasonably well be
514
described by their geometric structure and this structure can be
approximated by a polyhedra (skyscrapers, terrace houses), the
procedures of the system first are developed for polyhedral
objects in single images. It is also possible to extend the
interpretation process for buildings, consisting not only of
planes but of cylindrical surfaces or cones (castles, churches,
congress halls).
There are two possibilities for deriving the 3D information of
an object. The classical approach first applies point or edge
detection and possibly also grouping procedures to get a
symbolic description of the images (Fig. 1.a). This description
is either based on pixels, attributes and/or relations. Matching
the image descriptions of several images and determining the
object points in space by a spatial intersection allows a 3D
reconstruction of the object. The other approach applies edge
detection and grouping procedures in order to get sketches of
the object (Fig. 1b). A reasoning process then infers from the
2D features of the sketches to single 3D models of the object
which can easier be matched in space in order to get a 3D
description of the object. The advantage of this aproach is that
the problems involved in the step "matching in 2D" are
avoided, however replaced by the additional interpretation step
of sketches. Thus, in contrast to most existing approaches, the
matching process will take place in object space after the 3D
interpretation of the individual images.
Researchers have used various techniques for interpreting
sketches or line drawings. In their Mosaic image understanding
system HERMAN and KANADE (1984) implemented several
strategies to extract 3D information about houses from single
aerial images by exploiting the special structure of the houses
(vertical lines, flat roofs) and the special orientation of aerial
images. MULGAONKAR and SHAPIRO (1985) have
presented a PROLOG-based reasoning system which allowed
to interprete perspective line drawings. It contains a large
number of rules of the inverse perspectivity and of grouping
processes. The system is able to interpret even images basing
only on extracted line segments by means of these rules, but
seems to be inefficient for interpreting sketches of even rather
simple objects. HARALICK (1989) has collected a number of
Syml
3D d
Fig. |
a. Ck:
rules
from
SUC
cons
inter]
draw
and (
the p
form
mean
OT ev
edge:
The
geon
imag
and i
objec
imag
proc
sing]
numt
on th
are s
formi
basec
Char
cont:
syste: