You don’t need to be Jason Bourne to evade a computer that watches us through the images of a video camera. Just come across someone who looks like us, albeit vaguely, and the machine will probably follow another lead. However, this possibility of “fooling the” vigilante could come to an end after researchers from the University of Michigan (USA) have found theway to improve the accuracy of people tracking software by more than 30 percent. The trick: let the machine observe not only where an individual is going but also what he or she is doing.
“We have createda program that understands what activity a person or group is involved inand with this we can obtain much more robust and stable monitoring results “, explains Silvio Savarese, professor of electrical engineering, who assures that this is a completely new way of solving the monitoring problem. In fact, the new software recognizes such interactions as if a group of people are walking together or separately, if they are standing in a queue, if they cross the street, etc.
“The movements of an individual provide information about their interactions and the interactions can predict the future behavior of an individual,” clarify the researchers. For example, when two people are apparently walking together and talking, the computer can “connect their tracks.” Savarese and his colleagues taught the computer to “think like a human” by introducing videos of subjects with labeled behaviors, to contribute to pattern recognition based on past experience. “Our method reduces computational complexity andmakes it possible for the machine to infer what a person will do based on their activities as an individual, their interactions with other individuals, and behavior within larger groups“Savarese said.
The new software could be used in applications ranging from vision in robots to observing assemblages of wild animals or detecting suspicious and even criminal activity in a crowd. In addition, Savarese has a driving-related app in mind that would help drivers keep an eye on pedestrians by activating an alarm, or even the brakes, if a pedestrian takes an unexpected step on the road.
