onenten
inhalten.
1, erfolgt
lanalyse,
Objekte
eigt, wie
'eduziert
r die Be-
rund ge-
inschärfe
Begren-
Eignung
it diesen
ich ihrer
en Simu-
1. INTRODUCTION
Today real-time mapping concentrates on two to-
pics: the first topic addresses the combination of
data collection sensors like positioning and ima-
ging sensors for digital mapping. The second to-
pic is concerned with the development of mapping
procedures, i.e., methods and algorithms for the
identification and location of objects which have
to be captured and stored in a database. Because
of the real-time aspect of mapping simple but re-
liable procedures are required. With regard to
this second topic practice looks quite different. A
mapping equipment enables the user to identify
and locate objects in an absolute reference frame.
The automatic identification of objects is still a
hard problem of vision research.
Real-time classification procedures are most suc-
cessful in very restricted industrial scenes. In this
case the number of object classes is very small.
By thresholding the images the objects are sepa-
rated from the background. The extraction of
some simple features often is sufficient for the
identification of the objects. Significantly more
demanding are outdoor scenes, where in general
a lot of objects are present and the scene is com-
plex.
This paper focuses on the identification of sim-
ple objects in outdoor scenes. As an example we
use traffic signs and assume that the objects of
interest are located in a more or less natural en-
vironment. Such a scene might be captured by
a Highway Inventory System (Schwarz, 1992) or
similar surveying vans (Novak, 1990). For the ex-
periments in this paper we use a short sequence of
colour images. The images are taken by a stan-
dard video camera, i.e., the full frames (25 Hz)
are composed by the odd and even fields of the
half frames. Thus a full frame is the smallest mo-
tion unit in which information about the scene is
captured and, in addition, the effects of the mo-
vement of the car are represented.
The geometric structure or shape of the objects
of interest is simple. They can be modelled by a
plane which is spatially limited, for example, by
triangular, rectangular or circular border lines.
Traffic signs are typical objects of this class. For
the recognition the colour of those objects can be
expected to be an important clue. The use of a
small number of different colours like red, yellow,
blue, black and white keeps the discrimination
between objects relatively easy.
In developing a procedure for the identification
of objects from an image sequence some specific
problems have to be taken into account. In gene-
ral, the size of an object in the image is unknown.
The projection of the object leads to a perspective
distortion of the border line. Thirdly the move-
ment of the video camera during the exposure is
the reason for motion blur.
An important point for solving the mapping task
by an automatic procedure is the efficient use of
different information sources like the colour, mo-
tion and contour of the object. Simple algorithms
are required to achieve real-time capabilities. The
procedure we propose for the recognition of sim-
ple objects consists of the following steps. First
regions of interest, i.e. regions which may con-
tain the unknown objects, have to be detected.
Although in principle colour, motion and contour
may contribute to solve this task only motion is
used in this step. The result of the motion seg-
mentation is a displacement vector or displace-
ment field of this region. The displacement field
can be used to restore the image with the aim
of eliminating the effect of motion blur. The re-
stored image is well prepared for the extraction of
the border line of the object by which the location
of the object in the image is determined. Because
of the 3D to 2D projection of the object its bor-
der line is distorted. In consequence, for recogni-
tion it is advantageous to extract affine-invariant
features from this contour. The last step of the
procedure is a maximum likelihood classification.
The affine-invariant features can be used to deter-
mine the most likely object class for the unknown
localized object. In the following these main steps
of the analysis are discussed further. Concerning
the interpretation the suitability and the separa-
bility of the object features is of special interest.
Therefore, this aspect will be analysed in more
detail.
2. DETERMINATION OF THE
REGIONS OF INTEREST
A first step in the recognition of an object is to
identify a region in the image in which the sear-
ched object is supposed to appear. This region
is called the region of interest. In the case of
461