New video analytic software can detect violent behavior

Published 27 October 2006

University of Texas scientists make a leap with software trained to distinguish between a hug and a push, or a handshake and a knifing; system is so far 80 percent accurate, and commercial applications could be available within a year; some see future in smart television searches, too

Among the most exciting new technologies in the homeland security field is video analytics, the ability of surveillance systems to pick up and learn the characteristics of suspicious behavior. This saves security employees the trouble of having to closely monitor large banks of screens at once because the system automatically notifies them if it spots anything out of the ordinary. An airport camera, for instance, could be trained to alert officials anytime a van was parked in a certain area for more than five minutes, or anytime a person crossed a certain boundary line inside a secure facility. These systems are effective, but they are “limited to spotting unusual visual patterns and cannot pick out specific types of activity,” New Scientist reported. A system under development at the University of Texas at Austin (UT) hopes to change that with an intriguing new technology that can actually tell the difference between friendly and violent behavior.

The software involved begins with the simple task of accurately identifying the form of a human body. The advance is with its ability to recognize patterns in how two humans interact with one another — a push versus a hug, for instance. Of course, scientists had to confront the ambiguities of human relations head-on. A person offering a piece of gum strongly resembles one pointing a knife, and so “semantic analysis” is used to build up a profile for each type of behaviour. So, if two people approach one another with arms outstretched, the computer system looks to see if their hands move in synchrony. If they do, the computer recognzies the motions as a handshake. If they don’t, it recognizes it as a dangerous situation, perhaps a robbery or a fight.

Recent tests of the system found it 80 percent accurate at identifying a lengthy series of commonplace and dangerous activities. A commercial version could be implemented within the next few years, but already some are finding applications beyond homeland security. If loaded onto a smart television, for instance, one could very easily use the tecnology to search for fight scenes, or romantic ones.

-read more in this New Scientist report