Decoding the Subtle Dance of Ordinary Movements By ANNE EISENBERG Published: October 9, 2003 COMPUTERS are becoming sophisticated enough to identify people by the patterns of their voices, fingerprints and irises, but they cannot yet distinguish individuals by the subtle quirks in the way they move. That may change, though. Many researchers are working on ways to make computer programs better at spotting tiny expressive qualities in gait and gesture. Automatic recognition of these fine gradations in movement could lead to a host of uses, from improved security programs and earlier diagnosis of movement disorders to more lifelike computer animation. At New York University, Christoph Breg- ler, an assistant professor of computer science at the Courant Institute of Mathematics, and two colleagues have received a National Science Foundation grant to train computers to recognize the subtle movements people make. Dr. Bregler's approach combines the latest high-speed video technology with the principles of a Slovak dance theorist who flourished long before the age of computers. For the technology in his project, he has equipped his university laboratory with a dance floor and 10 high-resolution infrared cameras that can run at up to 1,000 frames a second. He records movement using a technique - one more familiar on a "Star Wars'' back lot than at the Courant Institute - called motion capture. Motion capture is a workhorse for animated films, video games and special effects. In one version of the technology, high-speed infrared cameras track the position of reflective markers pasted to bodysuits worn by performers. This information becomes the basis for an animated or computer-generated version of the same movement. The technology has limits, but Dr. Breg- ler, who has collaborated with movie studios, hopes to extend them. "Right now, the technology misses many ephemeral aspects of shape change," he said, qualities that artists must often painstakingly add. "Motion capture doesn't get at the individual expressiveness of how each person moves. The best animation is still handcrafted by animators." To add this missing ingredient to motion capture, Dr. Bregler has turned to the work of Rudolf von Laban, a dancer and choreographer born in 1879. Among his many innovations, Laban created a notation system known as Laban Movement Analysis to characterize fine shades of movements. Now, based on work with dancers and experts in Laban analysis in his lab, Dr. Breg- ler is assembling a databank of motions, translating them into mathematics and writing programs so that computers can recognize them. Edward Warburton, the director of the dance education program at N.Y.U., and Peggy Hockney, an expert in Laban Movement Analysis, are the other principal investigators on the National Science Foundation project. Dr. Warburton said Laban analysis was well suited to refining the broad strokes of motion capture. "Laban is concerned not just with the obvious stuff of where people are, but with the nuances of how they got there," he said. Laban's system addresses the difference, for instance, between waving cheerfully or sadly, or the many differences in touch. "There is the gentleness of touching someone as though your hands were feathers," Dr. Warburton said, as well as a touch strong enough to push a car. Human experts will be needed to categorize these subtleties of motion, but not for long, Dr. Bregler said. "In the beginning, we need people to do the identifying," he said. "But then the computer should be able to find these movements automatically." He wants the movements incorporated into his computer models not only so more realistic animation can be created, but also so that, for example, a computer program could analyze a patient's gait to assess the effect a drug was having on someone receiving therapy for Parkinson's disease. Other researchers inspired by Laban's work have added his insights to their computational models of movement. Norman Badler, an associate dean at the University of Pennsylvania and director of the university's Center for Human Modeling and Simulation, has been applying Laban's theories to distinguish between a threatening gesture and one of surrender, for example. "It's been the goal of my life to computerize Laban motion analysis," he said. "Laban analysis is a way of capturing fleeting qualities of movement, rather than solely their form." The qualities that Laban described are essential to representing motion realistically in animation, he said. Neglecting them leads to animated movements that look flat and robotic. Alex Vasilescu, a research scientist at the N.Y.U. Media Research Lab, , is also interested in how small inflections can be computerized. "I can pick out my dad from far away by how he moves," she said. "I think it's going to be possible to teach a computer to do this, too." Ms. Vasilescu hopes to identify what she calls a person's "motion signature," an individual style of movement that is as distinctive as a fingerprint. In a series of experiments using motion capture, she videotaped people as they were moving, extracting a mathematical description of what she calls their general movement style. Then she predicted how a person with a particular style would perform a different set of motions - say, climbing stairs. Finally, she compared the prediction with a motion-capture version of the person actually ascending or descending a staircase. So far, she said, the two versions have matched. "I've extracted a motion signature of how different people move regardless of the action - for instance, walking or climbing," she said. Dr. Michael Cohen, a senior researcher at Microsoft who has worked in computer graphics for 20 years, said that motion capture would benefit from the new research. "Motion capture is good at large motions," he said. "But fine subtleties? Certainly not yet." SOURCE: The New York Times http://tinyurl.com/qcwt * * * ---------------------------------------------------------------------- To sign-off Parkinsn send a message to: mailto:[log in to unmask] In the body of the message put: signoff parkinsn