Page 31 - Dynamic Vision for Perception and Control of Motion

1.5 What Type of Vision System Is Most Adequate? 15

A few years ago, many systems were investigated with single processors de-
voted to single pixels (Connection Machine [Hillis 1985, 1992], Content-
Addressable Associative Parallel Processors (CAAPP) [Scudder, Weems 1990] and
others). The trend now clearly is toward more coarsely granulated parallel architec-
tures. Since a single microprocessor on the market at the turn of the century is ca-
9
pable of performing about 10 instructions per second, this means in excess of
2000 instructions per pixel of a 770 × 525 pixel image. Of course, this should not
be confused with information processing operations. For the year 2010, general-
11
purpose PC processors are expected to have a performance level of about 10 in-
structions per second.
On the other hand, the communication bandwidths of single channels will be so
high, that several image matrices may be transferred at a sufficiently high rate to
allow smooth recognition and control of motion processes. (One should refrain
from video norms, presently dominating the discussion, once imaging sensors with
digital output are in wide use.) Therefore, there is no need for more elaborate data
processing on the imaging chip except for ensuring sufficiently high intensity dy-
namics. Technical systems do not have the bandwidth problems, which may have
forced biological systems to do extensive data preprocessing near the retina (from
120 million light sensitive elements in the retina to 1.2 million nerves leading to
the lateral geniculate nucleus in humans).
Interesting studies have been made at several research institutions which tried to
exploit analog data processing on silicon chips [Koch 1995]; future comparisons of
results will have to show whether the space needed on the chip for this purpose can
be justified by the advantages claimed.
The mainstream development today is driven by commercial TV for the sensors
and by personal computers and games for the processors. With an expected in-
crease in computing power of one order of magnitude every 4 to 5 years over the
next decade, real-time machine vision will be ready for a wide range of applica-
tions using conventional engineering methods as represented by the 4-D approach.
A few (maybe a dozen) of these processors will be sufficient for solving even
rather complex tasks like ground and air vehicle guidance; dual processors on a
single chip are just entering the market. It is the goal of this monograph to make
the basic methods needed available to a wide public for efficient information ex-
traction from huge data streams.

1.7 What Is Intelligence? A Practical (Ecological)
Definition

The sensors of complex autonomous biological or technical systems yield an enor-
mous data rate containing information about both the state of the vehicle body rela-
tive to the environment and about other objects or subjects in the environment. It is
the task of an intelligent information extraction (data interpretation) system to
quickly get rid of as many data as possible, however simultaneously, to retain all of
the essential information for the task to be solved. Essential information is geared
to task domains; however, complex systems like animals and autonomous vehicles