A radical new system that combines computer vision with the processing power of the human brain could soon allow intelligence community image analysts to examine surveillance footage at 10 times its normal speed.

Sajda’s portable prototype uses a laptop and an electroencephalogram (EEG) cap to pick out these neural signatures. Since the whole process depends on the very fast subconscious mind, the video footage can be run at 10 times its normal speed. At the end of the whole rapid-fire process, the identified images are picked out and re-sorted based on the strength of the “aha!” signals, allowing analysts to take a closer look only at what they recognized as most relevant.

Speeding up

“The cool thing about this is that we exploit the brain’s ability as a general visual processor, able to detect novel or interesting events or objects on the fly,” Sajda told HSToday. “The ultimate goal is to be able to process collected imagery when there aren’t enough analysts to look at it. Say, for example, 100 hours of imagery are collected per day per analyst, but an analyst can look at only 10 hours per day using conventional methods. They do a pretty good job of finding targets in those 10 hours of imagery, but what about the other 90 hours? So we speed up the video by a factor of 10 and use EEG to detect neural signatures of recognition events.”

The team says that it hopes to speed up the search process by 300 percent. If successfully developed, the device could provide immense aid to the intelligence community routinely, as well as speed up searches for suspects or potential terrorists in extremely crucial time-sensitive situations, such as the London bomb blasts.

“There’s a fundamental problem in current image analysis, particularly in defense-related applications, in that the number of images acquired each day vastly exceeds the ability to look at them all. Imagine the security guard confronted with multiple TV screens he has to monitor,” said Leif Finkel, professor of bioengineering at the University of Pennsylvania. “This is a very important contribution. The range of applications is huge — from analysis of medical imaging, to defense and security, to traffic monitoring.”

While the ultimate goal of the Defense Advanced Research Projects Agency (DARPA)-sponsored research is to create a device that will allow analysts to carry out extremely fast context-based searches, the added advantage is that the system will allow analysts to identify specific targets or unusual events; things that are quite difficult to do with existing computer vision systems.

“It would probably be possible for a single person to monitor five to 10 times the number of cameras that he or she could presently monitor,” said Steven Gordon, professor of information technology management at Babson College, Babson Park, Mass. “For many types of problems, such as a person putting an object into an unattended piece of luggage, a delay of a few minutes would not be critical between the action and the response. Computers are not very effective at identifying this type of activity, especially in the midst of a busy terminal, so the proposed technology could make it cost-effective to do this type of monitoring.”

No false positives

Since the system is designed to pick out potentially useful images to look at, experts say that false positives, if and when they occur, don’t really count.

“The visual system does not have to be ‘extremely’ reliable because this approach is used to triage the images, not to make the ultimate decisions,” said Misha Pavel, a professor of biomedical engineering at Oregon Health Sciences University.

However, Marco Graziano, CEO of Eptascape Inc., based in Mountain View, Calif., a company that specializes in intelligent video surveillance technology, was a bit skeptical of the entire approach.

“Accelerating the video analytics process is not the industry’s top priority, though DARPA may be interested in this kind of technology,” said Graziano. “Rather, the emphasis in the security industry is on automation that can relieve security guards from watching a steady stream of video. Conversely, the C3Vision approach would rely more heavily on monitoring personnel.”

The overall view seems to be that it will probably take some more time before robust artificial recognition systems are successfully put in place. “For the forseeable future, human-assisted recognition systems are probably the best option,” agreed Finkel.

Sajda’s team is currently working on improving the device’s ability to consistently detect EEG signals associated with interesting images. They aim to demonstrate the device to DARPA by the end of the summer.HST