The following is from EE Times, published by CMP Media. Primitive
machine vision is becoming a little more practical.
Posted: 11:45 p.m., EDT, 7/24/98
CMOS sensors open industry's eyes to new possibilities
By Yoshiko Hara
TOKYO - The advent of low-cost CMOS sensors modeled on the workings of
the human eye is opening new vistas in consumer-oriented imaging
applications spanning the gamut from diversions fit for videogames to
practical networks equipped with machine vision.
Recent demonstrations here by Mitsubishi Electric Corp. and Toshiba
Corp. show just how far-reaching the possibilities might be.
Mitsubishi's artificial retina, a version of which is already being
used in a low-cost camera for Nintendo Game Boy, will soon be pressed
into service to tell parking attendants when a spot opens up in the
lot. Toshiba's "motion processor," now in prototype form, recognizes
three-dimensional objects in real-time. You can use it to play
scissors, paper, rock with your computer.
The compact CMOS sensors that lie at the heart of these developments -
and their friendliness with other CMOS peripheral circuitry and with
standard manufacturing processes - hold out the promise of
applications aplenty at a potentially low cost.
"As the artificial-retina LSI filters image data, only a small amount
of data is output," said Kazuo Kyuma, manager of business promotions
for the device at Mitsubishi's system LSI division. "This is an
advantage, because [it means] a low-cost microprocessor such as a
16-bit microcontroller can handle the subsequent processing."
Mitsubishi sees the chip growing into what Kyuma called "a key device
that will be used widely everywhere as industry eyes." Starting this
month, the company will offer a development kit and library to help
users get their own systems up and running.
Toshiba, meanwhile, is planning a 2000 rollout for the motion
processor it demoed in prototype form. Built around a CMOS image
sensor, the device "can capture depth information in addition to the
[standard] two-dimensional data using only one sensor," said Miwako
Doi, chief researcher of human interface technology at the company's
communication and information-systems research laboratory.
"We are looking for appropriate applications. But in terms of
technology, we are ready to market," said Fumio Sugiyama, the lab's
general manager.
Bow to biology
Mitsubishi's artificial-retina LSI consists of a two-dimensional
variable-sensitivity photodetection cell array, a random-access
scanner for sensitivity control and a mutiplexer for output. It
executes programmable filtering on-chip, directly on the projected
images, just like a biological retina. Since there is no need for an
additional, high-performance processor, an image-recognition system
built around the chip can be compact, low in cost and power-thrifty.
As the first customer for the device, Nintendo used the artificial
retina in a camera for its portable Game Boy machine. The pocket
camera has resolution of only 128 x 128 pixels with four gray scales,
and sells for about $40. Players use it to capture images that can be
edited easily on the Game Boy.
"We used the device because of its low cost and low power
consumption," said a Nintendo spokeswoman. Introduced in February in
Japan and in June in the United States, Pocket Camera for Game Boy
sold more than 1 million units in the first two months, the
spokeswoman said.
In April, Mitsubishi moved the artificial-retina project group from
its Itami laboratories, where the LSI project spent its first eight
years, to its system LSI division in preparation for a full-fledged
market assault.
Timed with the debut of the Nintendo camera, Mitsubishi started volume
production of the 128 x 128-pixel retina LSI in a transparent package.
Capacity has reached 1 million units a month. Samples run about $11
each.
The new development kit consists of a sensor unit comprising lens, the
128 x 128-pixel artificial-retina LSI and a board on which an
analog-to-digital converter, FPGA controller and interface are
mounted. Working on a Windows 95 PC through the ISA bus, a user can
evaluate all functions of the artificial retina. A development library
offered simultaneously with the kit makes it possible to program in C
or C++. The kit and the library are priced at about $700 each.
changing the control-signal pattern applied on the scanner section,
the retina chip can output various images without using special logic
circuitry. It can handle simple image capturing, edge extraction,
pattern matching and random access, in which it opens a small window
and outputs only the data inside the window.
Using two 32 x 32-pixel modules, Mitsubishi researchers have completed
a prototype of a three-dimensional motion-detecting system as well.
Moreover, developers can leverage the chip's low cost, hooking
hundreds of sensor cameras into one network. Mitsubishi is
codeveloping with its affiliate company a parking control system using
128 sensors that, in tandem, report occupancy of the lot. Since each
retina chip has a recognition function, only the cameras that
recognize a status change send data, so a simple network can be
constructed even with a vast number of sensors, according to Kyuma.
The system will be in practical operation within the year, the company
said.
In addition to the initial 128 x 128-pixel LSI, Mitsubishi is
developing a chip with the same resolution in a ceramic package for
use in tough conditions. It is also working on a 32 x 32-pixel LSI and
a 352 x 288-pixel CIF (Common Intermediate Format) chip. The latter
two will sample by the end of the year, and volume production will
begin in mid-1999. Color sensors are also under development with
several partners.
"The LSIs are fabricated on existing lines, the 128 x 128 device on a
0.8-micron process and the coming CIF LSI on a 0.5-micron process,"
said Kyuma. "No additional investment is needed for these existing
lines."
Three in one
For its part, Toshiba's prototype motion processor enables real-time
recognition of a moving three-dimensional object with one CMOS sensor.
Using the motion processor as an interface for a PC, Toshiba
demonstrated a scissors-paper-rock game played to the accompaniment of
a synthesized percussion instrument on the computer.
In the demo, the motion processor accepted gesture input. For example,
the beat of the music can be controlled by swinging the hand close to
the processor, orchestra-conductor-style.
Separating an object from its background is an essential procedure in
recognizing a 3-D object. Conventionally, an object could be set apart
only when it was on a simple background or had special marks affixed
to make it stand out.
But Toshiba uses a different method. Its motion processor has eight
light-emitting diodes set around a lens that covers both the CMOS
image sensor and an LSI which controls synchronized light emission and
detection. The LEDs emit 940-nm light and the CMOS sensor detects
reflected light. The difference in the intensity of reflected light
enables the processor to select the object from its background and
gives depth information. To eliminate the interference of light,
detection is synchronized with the LEDs' emission. The time control of
emission and detection is the key technology, and Toshiba declined to
disclose details because a patent is pending.
Researchers used a 64 x 64-pixel CMOS image sensor for data transfer
efficiency. The motion processor detects objects at 30 to 90 cm and
sends matrix data of X, Y and Z coordinates to the PC at 30 to 50
frames/second; 3-D recognition is done at the PC with dedicated
software.
The aim is a compact product that will sell for about $70, said lab
chief Sugiyama.
[morenews.gif]
Click Here to Vist CMPNET Click Here to Vist CMPNET
_______________________________________________________________
[USEMAP]
All material on this site Copyright (c) 1998 CMP Media Inc. All rights
reserved.
This archive was generated by hypermail 2b29 : Sun Dec 02 2012 - 01:30:04 PST