Hardware Reference
In-Depth Information
Challenges of Identifying Physical Tokens
Designer Durrell Bishop's marble telephone answering
machine is an excellent example of the challenges of identi-
fying physical tokens. With every new message the machine
receives, it drops a marble into a tray on the front of the
machine. The listener hears the messages played back by
placing a marble on the machine's “play” tray. Messages are
erased and the marbles are recycled when they are dropped
back into the machine's hopper. Marbles become physical
tokens representing the messages, making it very easy to
tell at a glance how many messages there are.
with sticky-backed copper tape. When I went to Apple and
worked with Jonathan Cohen, we built a properly hacked
version for the Mac with networked barcodes.
Later, again with Jonathan but this time at Interval Research,
we used the Dallas ID chips.
Color by itself isn't enough to give you identity in most
cases, but there are ways in which you can design a system
to use color as a marker of physical identity. However, it
has its limitations. In order to tell the marbles apart, Bishop
could have used color recognition to read the marbles, but
that would limit the design in at least two ways. First, there
would be no way to tell the difference between multiple
marbles of the same color. If, for example, he wanted to use
color to identify the different people who received messages
on the same answering machine, there would then be no
way to tell the difference between multiple messages for
each person. Second, the system would be limited by the
number of colors between which the color recognition can
reliably differentiate.
Bishop tried many different methods to reliably identify and
categorize physical tokens representing the messages:
I first made a working version with a motor and large
screw (like a vending machine delivery mechanism), with
pieces of paper tickets hung on the screw, and had different
color gray levels on the back. When it got a new message,
the machine read the next gray before it rotated once and
dropped the ticket. It was a bit painful, so I bought beads and
stuffed resistors into the hole which was capped (soldered)
Shape and Pattern Recognition
Recognizing a color is relatively simple computationally;
but recognizing a physical object is more challenging. To
do this, you need to know the two-dimensional geometry
of the object from every angle so that you can compare
any view you get of the object.
Face Detection
If you've used a digital camera developed in the past five
years or so, chances are it's got a face-detection algorithm
built in. It'll put a rectangle around each human face and
attempt to focus on it. Face detection is a good example
of visual pattern recognition. The camera looks for a pre-
described pattern that describes, generically, a face. It has
a particular proportion of height to width: there are two
darker spots about one-third of the way from the top, a
second darker spot about two-thirds of the way down, and
so forth. Facial detection is not facial recognition—a face-
detection algorithm generally isn't looking specificially
enough at an image to tell you who the person is, just that
they have something that more or less resembles a face.
A computer can't actually “see” in three dimensions using
a single camera. The view it has of any object is just a
two-dimensional shadow. Furthermore, it has no way of
distinguishing one object from another in the camera
view without some visual reference. The computer has
no concept of a physical object. It can only compare and
match patterns. It can rotate the view, stretch it, and do
all kinds of mathematical transformations on the pixel
array, but the computer doesn't understand an object as a
discrete entity the same way a human does.
OpenCV has patterns for facial detection that are very
simple to use. The following project shows you how to
detect faces.
X
