Surveillance cameras can serve as an effective deterrent to crime, provided security personnel actually watch the video streams from the cameras. But having the staff resources to monitor dozens of cameras at once is beyond the reach of many organizations.
With behavior-recognition software, however, it all becomes possible. By layering video-analytics software over an existing closed circuit TV (CCTV) system, law enforcement can use video as a preventive measure because officers get alerted to crimes as they occur, instead of monitoring dozens of video screens at the same time or just using the video as forensic evidence.
The technology is being used successfully at Johns Hopkins University in Baltimore, where a runaway crime rate was significantly reduced. At the Rapid City, S.D., Regional Airport, the technology effectively replaced Transportation Security Administration personnel in some areas.
The software can be configured to recognize people's movements, such as someone running, and changes in landscape, such as a package left behind, and then alert appropriate personnel. A behavior-recognition system can free up staff to take on other tasks until there is an alert, or allow a smaller staff to monitor dozens of cameras.
"It is programmed to recognize a wide variety of things based on what you, as a law enforcement person, would care about," said Nik Gagvani, chief technology officer at Cernium, which manufactures a behavior recognition system called Perceptrak. "You can customize it to create what we call a 'behavior profile.' It would just report those behaviors to you."
A behavior-recognition module can be programmed to recognize up to 18 different behavioral patterns, including slow-moving vehicles; erratic movement; a person falling down; an abandoned object; or a number of people converging. The module can be programmed to assign a score to each alert based on the significance of the behavior.
The score is influenced by factors such as time of day and others programmed into the system. All alerts are then stored automatically, giving investigators forensic evidence.
How It Works
"If you think of video, it's really a series of images one after the other over time," Gagvani said. "If you look at any pair of successive images, and you subtract or find the difference of these images, you would see where it's changed over time."
The analysis is done by looking at how the pixels in the video change over time. The software is configured to recognize certain changes that correspond to movement.
It can also recognize, through a process called segmentation, if something has been left behind, such as a backpack that isn't part of the landscape.
A video scene includes ambient props such as trees, buildings and other permanent structures. That's all considered background. Anything that moves through the scene is considered foreground. Segmentation is the process of separating the background from the foreground. If a bag was left in a scene, for instance, it would not be part of the background and would trigger an alert from the system.
"Now the question is, 'Do the differences mean anything from a security point of view?'" Gagvani said. "That's what behavior recognition is all about. It's not just about what is referred to as motion detection, which has been around for a while, but now trying to qualify that motion and make sense of it. It gives you information you can act on. You don't have to look at the video until one of these things you care about is discovered by the behavior recognition module."
Then it's up to law enforcement personnel to look at the video feed and assess whether action needs to be taken.
"Basically the technology is trying to be the unblinking eye, so to speak," Gagvani said. "It's not as sophisticated as a human, but it doesn't experience fatigue like a human would. What it does not do is tell you that there is a threat. That qualification has to be made by a human. It is an assistant, and given the magnitude and widespread deployment of CCTV and surveillance technology these days, it's simply an intractable problem to look at all that video manually."
While the system doesn't recognize a certain individual or identify a threat, it will recognize behavior it has been programmed to recognize and then alert someone. It's resilient to rain, wind and blowing debris but doesn't work in dark settings. It doesn't work well in extremely crowded conditions, when there's just too much movement.
"In airports, it's used for a very specific purpose: to monitor the exit lanes where you expect people to be going from the gate area to the terminal," Gagvani said, referring to the deployment of the technology at the Rapid City Regional Airport. "This technology works like a champ when you define specifically what you care about and not have it do 20 different things."
The system can be expanded to yield more or fewer alerts by being configured to respond to more or fewer behaviors or movements. "The trick is to balance false alarms versus true alarms," Gagvani said. "And to keep the false alarm rate as low as possible because if this is alarming all the time, you might as well just look at the video all the time."
Keeping the false alarms at fewer than 10 percent of the total alerts is the goal, he said. "A well run behavior recognition system should give you no more than 100 alerts a day."
A system includes the cameras, the cabling that brings the video back to the security center or wherever it's being viewed and the video processors.
A Safer Campus
At Johns Hopkins University, a state-of-the-art communications center houses the brains of 89 "smart" cameras positioned strategically around the campus, and the 14-member staff that monitor the cameras.
"They installed a fairly elaborate video surveillance and emergency communications system and put our video analytics functionality into it," said Cernium CEO Craig Chambers. "It's now their primary source of dispatching officers on campus based on potential threats or interesting situations."
When the system recognizes behavior that it's been programmed to read as unusual, it alerts a staff member by framing the scene with a yellow rectangle. Staff then decides whether to send campus police or not.
Since its deployment in March 2005, the system has alerted staff to numerous thefts in progress, vandalism, minor traffic accidents and helped catch an armed robber. Campus bike thefts dropped from 25, during the fall 2005 semester, to three in fall 2006. The total number of crimes reported on campus in 2003 was 536. That number grew to 703 in 2004 then dropped to 279 after the installation of the new system in 2005.
Though the goal at the 140-acre Johns Hopkins University campus is to reduce crime overall by honing in on specific behaviors, the Perceptrak system deployed at the Rapid City Regional Airport has one function: to "watch" over the safety checkpoint areas and make sure nobody bypasses security.
"All we're looking for is wrong-way motion, which is a simple application, but it's a tremendous labor savings," said Mason Short, executive director of the Rapid City Regional Airport. "There's nothing more irritating for me than to walk out into the checkpoint and see screeners sitting on their butts doing nothing other than making sure somebody doesn't walk through. It's a waste of labor and manpower."
Two of the three cameras used at the airport are configured to sense wrong-way motion, and the third is a monitor that allows for Internet access from virtually anywhere. It's used to review incidents.
The setup cost approximately $100,000, but the airport spent an additional $80,000 on construction of a physical structure in the checkpoint area to replace the glass barriers. The glass inhibited the behavior recognition system because the camera responded to reflections from the glass. Short said the setup results in a savings in checkpoint personnel of between $100,000 and $120,000 a year.
"It's a narrow passageway where we put everybody through past the checkpoint," Short said. "There's a pre-alarm area so that if people are just mistakenly walking the wrong way, we've got a camera that will sense the wrong-way motion and scare them out of their pants with a screaming, 'Stop! Halt! Back up.' If they proceed past the pre-alarm area, they go into the full alarm area, where there's another camera doing behavior recognition."
If a full audible alarm goes off, airport security personnel and local police respond to the call. The system also stores any alerts, allowing investigators to go back and review incidents or submit video as evidence as was the case in following a 2006 airport incident.
"We had an incident last year where an individual ran through the checkpoint to go back to an aircraft," Short said. "He set off the pre-alarm, and it didn't faze him. Then he fully alarmed the system, became belligerent when law enforcement showed up, and he was actually incarcerated for violating federal airport security regulations.
"The key was being able to bring this piece of evidence, this video evidence, to the judge and say, 'This is what he did.' Without that, we wouldn't have had a case."