USING PERCEPTUAL GROUPING FOR ROAD RECOGNITION
G. Forlani*, E. Malinverni*, C. Nardinocchi**
*DIIAR - Politecnico di Milano , Piazza Leonardo da Vinci, 32- 20133 Milano, Italy.
**DScMtTe - Università di Ancona, Brecce Bianche, 60100 Ancona, Italy.
gianfra(gipmtf].topo.polimi.it
eva @ipmtfl.topo.polimi.it
carla@anvax1.unian.it
Commission III, Working Group 3
KEY WORDS: Vision Sciences, Classification, Recognition, Feature Extraction.
ABSTRACT:
Automatic localization and identification of cartographic object from aerial and satellite images has gained an increasing attention
in photogrammetry. The approaches for automatic extraction of man made objects may be grouped into two broad categories:
semi-automated methods and fully automatic systems. Here an automatic system oriented to road recognition is presented.The
system is based on a three stage procedure: image segmentation by feature extraction, perceptual organization of the geometric
attributes of the features and object recognition based on an implicit knowledge base representation.
1. INTRODUCTION
On the way towards automatic mapping and GIS data
acquisition and update, automatic localization and
identification of cartographic object from aerial and satellite
images has gained an increasing attention in
photogrammetry, while DTM generation and thematic
classification have already reached a high degree of
reliability, the automatic extraction of man made objects,
which are of major interest in applications, is still an
unresolved issue. The approaches which are currently
pursued may be grouped into two broad categories: semi-
automated methods rely on the intervention of human
operators to provide either the initial input to the system
(e.g. seed points) or to help the system to bridge loopholes
or ambiguities (Vosselman & de Knecht 1995; Gruen et al.,
1994); fully automatic systems on the contrary should be
able to address the whole complexity of the task: therefore
they need to implement a more refined strategy based on a
set of rules or assumptions which constitute the knowledge
base of the system (Barzhoar & Cooper, 1995; Steger et al.,
1995). Apart from the definition of a convenient and
effective strategy, there is still a lot to improve in the
fundamental stage of feature extraction, since many of the
algorithmic problems arise from the poor quality of the
extracted edges. At present, only semi-automated systems
represent a good compromise between the speed of the
procedure and the time and the commitment required to the
operator. Still, research should aim towards increasing
automation, since, apart from cost reasons, most of the
interaction required would be anyway too repetitive and
may prove, if the process stops too often, less appealing than
manual plotting by the operator.
We are working on the development of an automatic image
analysis system oriented to recognition of cartographic
objects. The system is based on a three stage procedure:
e image segmentation by feature extraction
e perceptual organization of the geometric attributes of the
features
e object recognition based on an implicit knowledge base
representation.
In its current stage of development, only road recognition is
available and therefore that's what we are going to talk
about in the following.
Mm anms
202
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
2. IMAGE PROCESSING
2.1 Noise reduction
Before any image segmentation is performed, it is necessary
to reduce the amount of image noise. There are two main
types of noise in images: impulse noise and distributed
noise. The former affects the gray value only in some pixel
in the image, but to a large extent: it may be termed as a
gross error. As such, its effect on the edges is only local and
may be neglected. The latter affects all pixels and may be
assumed to be randomly distribuited, therefore appropriate
filtering is required. A trade off is to be found between the
edge smoothing implied by all low pass filters and the noise
reduction. Linear filters give a marked smoothing, so non-
linear filter are preferred in edge detection. The most
effective, but for the median filter, ideal for treating impulse
noise, are the Edge Preserving Smoothing (EPS) and the
Conditional Averaging Filter (CAF). In images with a small
noise content CAF performs better than EPS, since it
maintain more details and shows a more accurate edge
localization, while both are equivalent otherwise. We used
therefore CAF in our preprocessing stage, setting its
threshold by visual inspection.
2.2 Image segmentation
Segmentation groups the image pixels in regions satisfying a
given criterium; it may be based on texture or edge
properties. Here the latter approach is used, based on two
gradient characterists: magnitude and direction. In order to
detect linear features in the image and to ease the road
recognition stage, continuous lines will be approximated by
a sequence of line segments.
To select edge pixels a threshold must be introduced on
gradient magnitudes. Moreover, as long as the gradient
orientation remains pretty much constant along contiguous
pixels, they belong to an edge which is a line segment.
There are many alternatives in the way the gradient vector
may be computed and different threshold may be fixed for
its magnitude. The segmentation output therefore will be
dramatically affected by this choices, either making life
easy for the algorithms or preventing them from getting any
acceptable outcome. We based our image preprocessing on
the idea of carrying all information until we are in the
condition t:
In the gra
smoothing,
fine detail,
2x2 mask
response W
Figure
For each
computed
threshold,
following,
will alway
proceed w
enjoying s
because tl
space of !
equally sp
2).
Fig
2.3 Featu
We look
lines. Thi
following,
Brown, 1
(Burns et
is the fol
where the
straight li
gravity ce
to gradiet
determine
straight li
The segm
(either H.
which sor
squares a
small nur
as a weig
The choic
too many
large part
We found
the actual
At the
represent: