Imagine you are a terrorist being hunted by the authorities, and that you have disguised yourself with heavy makeup, wigs, or even a facelift. Is your disguise likely to go undetected? Not with some of the newest developments in antiterrorism technology, developed by U of T Ph.D. candidate Alex Vasilescu.
Her research into face recognition and human movement analysis, based on computer vision technologies, has resulted in software programs that have so much potential for filmmaking, biometrics, and security, that her research is funded by the U.S. Department of Defense. They have both garnered her U.S. patents, and won her MIT’s prestigious Technology Review 100 Award, which recognizes the top 100 young innovators each year.
“Face recognition is a difficult problem for computers,” Vasilescu said. The images contain too many complex “factors relating to scene structure, illumination, and imaging,” which tend to confuse a computer’s recognition system.
Attempting to unscramble the complications of visual images, Vasilescu used multilinear algebra, which provides a rigorous mathematical framework to unravel the entangled factors in an image ensemble.
“A TensorFaces representation has several advantages” over conventional face recognition systems, said Vasilescu, since they took into consideration the “different expressions, head poses, and lighting conditions.” The resulting software separates the many interacting factors in images and makes for better recognition.
Vasilescu has also developed Human Motion Signatures, a program which is able to compute a person’s unique motion signature-the individual way in which somebody moves. A shot of a masked terrorist on film stores sufficient information that can later be used to recognize him. With sample motions from a group of subjects and a previous shot of the suspect, the software can synthesize animations of a certain individual, say, escaping. Vasilescu’s developments are not limited to applications in security.
“With motion capture samples of Charlie Chaplin’s walk, is it possible to synthesize other motions, say, climbing up stairs in his distinctive style?” Vasilescu is excited by the thought that “human motion signatures” could be extracted from sample actions as human handwriting is extracted from sample writing.
“Imagine a car equipped with an automatic face recognition system that recognizes and customizes itself to each of the people authorized to drive it,” said Vasilescu. “Imagine a house that recognizes you and unlocks the door, despite the fact that you’ve lost the house key.”
Vasilescu’s research has also led to the development of TensorTextures. This new software that she developed in July is able to display the three-dimensional appearance of fabrics under all kinds of realistic lighting and viewing conditions.
“A certain material might be beautiful, but a potential customer wouldn’t know that because the image gives a grossly incomplete sense of texture,” said Vasilescu. Again, similar to the methods she employs in face recognition, she unscrambles the complex interactions between lighting, viewpoint, and geometry. Thus, the visual effects of soft cotton, rich velvet, smooth leather or glittering silk are no longer limited to fluorescent store lighting. The full range of effects can be visualized at all possible angles, perspectives, and lighting conditions, ranging from brilliant chandeliers to dim stage lighting, track lights to candlelight. Vasilescu is hoping that the software will be introduced into the market in a year.
Vasilescu currently does research in the Media Research Lab at New York University. She was honoured on Sept. 24 and 25 at the Emerging Technologies Conference held at MIT for her innovations. More detailed descriptions and articles may be found at http://mrl.yu.edu/alex/.
