Imagine a security camera that can understand and analyze the scene it's capturing. The camera can spot someone who is leaving a piece of luggage behind at an airport and issue an alert. It can identify a car being sought in connection with a crime based on its color, shape, and dented left bumper, even if the license plate has been covered up. And it can spot drivers who are texting when they should be watching the road, potentially passing on this information to law enforcement officials that can then issue a ticket.
These are some of the advanced security techniques being hyped by Movidius, a chip maker specializing in artificial intelligence and computer vision. Its Myriad chip powered the spatial awareness of Google's early Project Tango devices and the sense and avoid features inside DJI's latest Phantom 4 drone. The company was recently acquired by Intel, and today announced that its Myriad chip is being added to devices Hikvision, one of the world's largest sellers of internet-connected security cameras.
Over the last five years, deep learning, an approach to artificial intelligence, has sparked incredible improvements in computer vision. It's this progress that has allowed drones to dodge obstacles and driverless cars to park themselves. But it's tough to bring those techniques to mobile devices, because they typically require a lot of processing power.
Movidius has designed its chips from the ground up to perform deep learning techniques without overwhelming the smaller processors and batteries you find on mobile devices. It's specifically focused on computer vision, dubbing its Myriad chip a VPU, or vision processing unit. The name plays off the typical CPU you would find in a personal computer, or the GPU, graphic processing unit, you would find in a PC built for high-end gaming.
Movidius claims that because a lot of the processing and image analysis can be done on the device, the amount of data that needs to be transmitted back to a central server or human operator is much smaller. An intelligent security camera might be able to only share short video clips when it detects a serious threat or identifies a problem that requires manual intervention. That would mean only short highlights are transmitted, instead of a continuous video feed.
The company also says that the computer vision enabled by its VPU lowers the rate of false positives. As an example it passed along the video above, which shows a Hikvision camera in action without the help of a Myriad chip. The camera can intelligently identify objects, like a car, and recognize when they enter a restricted area. But it also registers an alert when a nearby bush, blowing in the wind, briefly enters into the restricted area. It's on the lookout for any kind of motion.
Security cameras with this level of visual intelligence are both exciting and worrisome. Hikvision cites the case of a senior citizen who became lost and was located using the facial recognition built into its cameras. But a camera that can find a lost child or identify a criminal suspect can just as easily be used for invasive government surveillance, like the case of protesters targeted through facial recognition reported earlier this month.
A camera running the Myriad chip, by contrast, would be able to classify the objects it sees moving throughout the scene. A car would be something that should trip an alarm, while a bush would be something it can safely ignore. Future cat burglars take note: you may be able to outwit the world's most advanced security cameras if you dress up as a very convincing piece of shrubbery.