| ALERTING MECHANISMS IN ATTENTIVE SYSTEMS |
Giovanni Garibotto
e-mail:
Giovanni.Garibotto@elsag.it
Attentive Vision represents one of the dominant paradigms in applied Computer
Vision. Beside representing a natural extension of biological vision, it has
been proved to be
the most effective approach to solve real industrial problems. Looking for
interesting points and areas in the scene is a mixture of a bottom-up process
(driven by relevant perceptual
features) and a top-down scheme (driven by goals and tasks to be solved).
The paper will refer some recent practical experience of this approach in order
to establish a basis of discussion.
Two deeply different areas will be considered, namely that of reading
technologies and that of mobile robotics,
to demonstrate how close is the vision requirement no matter
how far could appear the application fields.
In both cases the main requirement is the hierarchical and progressive
identification of the interesting
fields where to look closely in. When reading or understanding a mail piece
(either a letter or a flat or a parcel) it is necessary to detect,
distinguish and recognize very
different features and items (a bar-code, a stamp or mail priorities, the
destination address, the address of the sender, advertisement messages,
possible order-mails, etc.). Actually the understanding of the mail information
is becoming far more
complex than just OCR reading. Context information plays a fundamental role in
this game,
both in telling apart the different meaning of the available information as
well as
in improving significantly the performance of the specific recognition task.
In mobile robotics the use of artificial visual landmarks (as traffic signs,
special geometric
marks and lanes on the navigation floor) has proved to be an engineering
efficient approach to autonomous navigation control. The possibility to use
existing natural features
already present in the scene to be detected and tracked in a sequence of
images is a great challenge to extend this autonomous guidance capability.
In both cases the identification of features to be tracked is a goal-driven task
which is
based on alerting mechanisms acting as a focus of attention for the following
Computer Vision task.
This can be highly conditioned by the a-priori knowledge of the expected
3D model (like the known geometry of the landmark,
its symmetry properties, its perspective and
invariant properties, if available).
There are still a lot of unresolved problems and open issues in this approach,
as far as
control mechanisms are concerned (centralised vs distributed), the most
appropriate use of visual models
(high level models or low-level features), the use of feedback in
multiple resolution schemes.
The paper will address some of these questions in the framework of applied
computer vision
with a series of concrete industrial examples, to stimulate a discussion
on this matter.
e-mail: hmp@ipvvis.unipv.it