By Ann Geracimos
THE WASHINGTON TIMES
stuff of science fiction is coming to life in the work of computer
scientists studying human gait patterns.
They are working on the hypothesis that
each person has a unique gait so that one day our so-called
signature motion will be as valuable as a fingerprint in charting
Facial-recognition systems are
in use and are comparatively well-advanced. Extending the idea to
include the whole body, though, raises a host of problems that the
mathematical formulas behind the relevant software have yet to
The challenges are immense
because they involve capturing subtle changes in a person's walk
that can be compromised by shadows, different walking surfaces and
what a person carries. However, if successful, the systems would be
valuable in security surveillance as well as important diagnostic
tools for physicians who deal with movement disorders.
"At the time we started doing this work,
homeland security wasn't an issue; the project only really started
as a project for computer animation and for medical applications,"
says Alex Vasilescu, a research scientist at New York University and
a computer science doctoral candidate at the University of Toronto.
Her work on facial recognition — known
in tech speak as "TensorFaces: Multilinear Tensor Decomposition of
Image Ensembles" — led Massachusetts Institute of Technology's
Technology Review last year to name her as one of the top 100 young
scientists in the country.
An NYU Web
site explains that the purpose of facial-recognition software is "to
enhance computers' ability to match multiple characteristics of a
face in ways that overcome vagaries of shading, angle or
expression." Gait patterns, formally called "human motion
signatures" by researchers, are based on being able similarly to
extract and analyze characteristics of movement patterns.
"I wanted to extract an individual's
personal style and the explicit manner in which they move in a way
that translates across different motions and is consistent," she
explains in a telephone interview. Her initial interest grew out of
a common observation that friends often recognize one another at a
distance by shape or movement alone before it's possible to see the
face. She wondered, she says, if she could teach a computer to
perform in the same way.
"At the time,
I was thinking of all these actors who have unique ways of moving,
and I wondered how to translate that into a signature gait or style.
You can capture this and sit there and play back the motion. If you
take that particular motion, you will ask, 'What is the signature
here, and how does that change when he does anything else, like
running up- or downhill — a signature motion that would reveal other
Next, Ms. Vasilescu
says, the researcher has to find how close a signature motion is to
normal so that a person who has gone for physical therapy, for
example, can have his progress charted using this system.
"People usually say that they feel
better after therapy, but that is a fuzzy word. Using this, you can
see improvement and a way of quantifying different treatment
methods," she says.
She knew she easily
could recognize someone familiar, but what happens, she wondered, if
you can't see the person's face in the dark and he or she is not
looking into a camera? "What is another modality to use based on the
way they move?"
Some work along these
lines had been explored by psychologists as far back as the 1960s,
she says. The magazine Scientific American, she recalls, had written
about how reflective markers placed on different joints allowed an
observer to trace the way dots were moving through a dark space.
That prompted the question of how a person could deduce what was
happening in that particular scene.
David Herrington of the Technical
Support Working Group, an interagency organization that funds
anti-terrorism research through the Department of Defense, praises
Ms. Vasilescu's work on solving basic mathematical problems involved
in such research but calls the technology immature. He says it might
take 10 years for human gait studies to prove themselves.
"Ten years in this business is an
eternity," notes Larry Davis, chairman of the department of computer
science at the University of Maryland, which has received research
money in the past from the Department of Defense's Advanced Research
Projects Agency (DARPA) for the Human Identification at a Distance
project. The study involved five universities around the country
interested in developing computer algorithms that could identify
people by how they walk.
measured by how well a program could compare a person's gait against
all the gaits in its database.
project ended last fall because it was in the same office as the
Total Information Awareness project, which was shut down, he says.
"It would have ended anyway," he adds. "There wasn't really enough
work done to come to any firm conclusion. It wasn't a technology
ready for incorporation into any surveillance [method]."
When gait research is perfected, its
uses for anti-terrorism surveillance will be invaluable, says a
former program manager at DARPA who asked that his name not be used.
He mentions the possibility of having software that would detect
whether a stranger walking into a facility is a frightened woman or
a terrorist with hidden explosives.
Entrepreneurs who have developed
computer software sophisticated enough to separate objects of
interest from the background for security surveillance purposes — a
step along the path to individual-recognition systems — include
ObjectVideo of Reston. The company's VEW, for video early warning,
software can be programmed specifically to send an alert only after
determining whether a movement or object constitutes a danger,
according to Alan Lipton, VideoObject's chief technology officer.
It, too, can be traced back to
DARPA-funded work on computer vision technology for defense
purposes. A customer can set the rules verbally by giving the system
what Mr. Lipton calls policy statements on what to watch out for. It
isn't motion or gait detection per se, but a way of detecting when
people are doing unusual things. When instructed, the system can
send an alarm, for instance, about a bag left unattended for a long