A camera developed by computer scientists at the University of California, Berkeley, would obscure, with an oval, the faces of people who appear on surveillance videos. These so-called respectful cameras, which are still in the research phase, could be used for day-to-day surveillance applications and would allow for the privacy oval to be removed from a given set of footage in the event of an investigation.
“Cameras are here to stay, and there’s no avoiding it,” says UC Berkeley computer scientist Ken Goldberg. “Let’s figure out new technology to make them less invasive.” According to a 2006 report prepared by the New York Civil Liberties Union, the number of publicly and privately owned video cameras in Lower Manhattan increased by a factor of five between 1998 and 2005, and several thousand cameras are in place in Greenwich Village and Soho alone. The United Kingdom, however, holds the record for video surveillance. In a report filed on Tuesday, the information commissioner there estimates that there are four million video-surveillance cameras in the United Kingdom–that’s one for every 14 people. Goldberg thinks of the respectful cameras as a compromise between advocates for privacy and those concerned about security.
In its current state of development, the camera is only able to obscure the faces of people who are wearing a marker, in the form of a yellow hat or a green vest. The camera system was developed by the National Science Foundation-funded Team for Research in Ubiquitous Secure Technologies, and it currently works in real time with Panasonic’s robotic security cameras operating at 10 frames per second and a resolution of 640-by-480-pixel videos. The researchers use a statistical classification approach called adaptive boosting to train the system to identify the marker in environments with a high degree of visual noise. But they also combined this classifier with a tracker, which takes into account the subject’s velocity, along with other interframe information. At a construction site where the researchers tested their camera with the vest, the system correctly identified the marker 93 percent of the time. Under more-uniform lighting conditions in their lab environment, they report 96 percent success at identifying the hat, even when two marked individuals cross paths.
The marker requirement is a trade-off, Goldberg admits, but he says that face-detection algorithms are simply not up to task for real-time operations in complex environments. “The idea is called structuring the environment,” he says. “If you’re willing to meet the system halfway and say, ‘I’ll help the computer,’ then that’s useful.” In areas with heavy surveillance, markers could be made available, just outside the camera’s view, to those who wish to maintain their privacy. In the future, Goldberg says, it may be possible to use a less conspicuous marker, like a button, particularly with systems of multiple cameras, which would be less susceptible to visual obstructions.
Gain the insight you need on robotics at EmTech MIT.