In these days of heightened security and precautions, surveillance cameras watching over us as we cross darkened parking lots or looking over our shoulders at airports may seem reassuring, but they’re only of use if someone is watching them. Researchers have found a way to give these cameras a rudimentary brain to keep an eye out, and the research is already been licensed to a New York company with an aim toward homeland security.From the University of Rochester:Smart Software Gives Surveillance Eyes a ‘Brain’
In these days of heightened security and precautions, surveillance cameras watching over us as we cross darkened parking lots or looking over our shoulders at airports may seem reassuring, but they’re only of use if someone is watching them. Researchers at the University of Rochester’s computer science laboratories have found a way to give these cameras a rudimentary brain to keep an eye out for us, and the research is already been licensed to a Rochester company with an aim toward homeland security.
“Compared to paying a human, computer time is cheap and getting cheaper,” says Randal Nelson, associate professor of computer science and creator of the software “brain”. “If we can get intelligent machines to stand in for people in observation tasks, we can achieve knowledge about our environment that would otherwise be unaffordable.”
Far from being an electronic “Big Brother,” the software would only focus on things for which it was trained to look?like a gun in an airport, or the absence of a piece of equipment in a lab. Nelson has even created a prototype system that helps a person find things around the house, such as where reading glasses were left.
Nelson set about experimenting with how to differentiate various objects in a simple black-and-white video image like that used in a typical surveillance camera. The software initially looks for changes that happen within the image, such as someone placing a cola can on a desk. The change in the image is immediately highlighted as the software begins trying to figure out if the change in the image is a new object in the scene, or the absence of an object that was there before. Using numerous methods, such as matching up background lines that were broken when the new object was set in front of them, the prototype system is accurate most of the time. It then takes an inventory of all the colors of the object so that an operator can ask the software to “zoom in on that red thing” and the software will comply, even though the soda can in question may be red and silver and overlaid with shadows.
The next step, however, is where Nelson’s software really shines. Nelson has been working for years on ways to get a computer to recognize an object on sight. He began this line of research over a decade ago as he wrote software to help a robot “shop”?picking out a single item, like a box of cereal, from several similar items. One of the tasks he recently gave his students was to set up a game where teams tried to “steal” objects from one another’s table while the tables were monitored by smart cameras. The students would find new ways to defeat the software, and consequently develop new upgrades to the system so it couldn’t be fooled again.
Though a six-month-old baby can distinguish different objects from different angles, getting a computer to do it is a Herculean task of processing, and more complicated still is identifying a simple object in a complicated natural setting like a room bustling with activity.
Unlike the baby, the software needs to be told a lot about an object before it’s able to discern it. Depending on how complex an object is, the software may need anywhere from one to 100 photos of the object from different angles. Something very simple, like a piece of paper, can be “grasped” by the program with a single picture; a soda can may take half a dozen, while a complex object like an ornate lamp may need many photographs taken from different angles to capture all its facets. With those images in mind, the software matches the new object it sees with its database of object to determine what the new object is.
The technology for this ‘smart camera’ has already been licensed to the local company PL E-Communications, LLC., which has plans to develop the technology to control video cameras for security applications. For instance, CEO Paul Simpson is looking into using linked cameras covering a wide area to exchange information about certain objects, be they suspicious packages in an airport or a suspicious truck driving through a city under military control. Even unmanned aerial reconnaissance drones like the Predator that made headlines during the current Iraqi war can use the technology to keep an eye on an area for days at a time, noting when and where objects move.
“We’re hoping to make this technology do things that were long thought impossible?making things more secure without the need to have a human operator on hand every second.” says Simpson.
Nelson and PL E-Communications were connected through the Center for Electronic Imaging Systems (CEIS), a NYSTAR-sponsored Center for Advanced Technology (CATs) devoted to promoting economic development in the greater Rochester region and New York State. CEIS develops and transfers technology from local universities to industry for commercialization, and by educating the next generation of leaders in the fields of electronic imaging and microelectronics design.