We are increasingly being watched by security cameras, and that’s no longer a figure of speech. Soon the cameras themselves will be watching—and figuring out if would-be thieves are shoplifting, terrorists are planting bombs, or drivers are texting behind the wheel. Machine-vision developer Movidius and security camera maker Hikvision have announced a new line of cameras (yet to be named) equipped with deep neural network AI that they claim can figure out if something sketchy is happening and send an alert in real time. The goal is to use security cameras to stop things before they happen, rather than archive footage for investigators to look through afterward.
AI-powered image analysis is already making its way into newer security camera systems—using the digital IP cameras that are beginning to take over from analog CCTV models. But the computing happens in the cloud, far from the scene of the potential crime and therefore with some delay. “There is no ability to proactively act on what you are capturing,” says Movidius CEO Remi El-Ouazzane. Movidius and China’s Hikvision (which supplies about a fifth of the global video surveillance market) claim that their devices represent the first time that the brains live in each camera. Movidius makes the bold claim that analyzing video on the device instead of sending it off to the cloud for processing could be 1,000 times faster and up to a million times more energy efficient. “We are living in a world where real-time means a few milliseconds,” says El-Ouazzane.
Hikvision’s cameras (and future models) will be equipped with the Movidius Myriad 2 vision processing unit, the same chip that DJI‘s new Phantom 4 and Mavic Pro drones use for autonomous navigation. Movidius (just acquired by Intel and engaged in a hush-hush collaboration with Google) has upped the game for mobile artificial intelligence—in this case, vision-based analysis. AI has been making its way into smartphones, taking advantage of the boost in mobile CPUs and graphics processors. Movidius has taken it further with Myriad 2, a separate chip dedicated to analyzing images using convolutional neural networks (CNN) that mimic how the brain processes vision. CNNs are en vogue—not because they are new, but because computers are finally powerful and efficient enough to run them. The Myriad 2 sips a mere 1 watt of power. “Movidius is kind of the leader in vision hardware today,” says Alberto Rizzoli, cofounder of Aipoly, a company whose eponymous iPhone app identifies objects for blind people.
Movidius will soon have competition, however. Earlier this month, mobile chip giant Qualcomm announced a reference design (the core components) for deep-learning cameras. Based on its Snapdragon 625 mobile-phone chip, the system will be available to camera makers by the end of the year, says Qualcomm. But Movidius does have a head start. “We have some other big stuff coming,” says El-Ouazzane, mentioning not only cameras but drones, augmented and virtual reality, and wearables.
Movidius claims that its technology could head off theft or terrorism by using Hikvision’s existing security camera footage to learn what suspicious actions look like. “The moment you can train a neural network to…reproduce the behavior of a theft,” says El-Ouazzane, “you may have the ability to, based on the behavior of an individual, prevent this theft from happening.” A pretty straightforward example is watching to see if someone hangs around or keeps coming back to a store, possibly casing it to plot a theft. Likewise, systems can detect if someone breaks in to an area or leaves a suspicious package that might be a bomb. Recognizing sketchy activity, the camera can send a real-time alert to a store clerk or security guard.
Trickier is spotting something like shoplifting. “I’ve seen examples online of people claiming to be able to detect shoplifting behaviors, like reaching out for a box and putting it in a pocket or inside a coat,” says Rizzoli. “But there is no research to back it up.” (Rizzoli isn’t directly affiliated with Movidius.) El-Ouazzane claims that, with enough training, neural networks will be able to learn more subtle activities. Putting AI in the cameras crowdsources the development of vision algorithms. Cameras will upload lessons learned to Hikvision’s servers in order to generate better algorithms, which are periodically pushed back down to each camera.
Much easier is detecting traffic violations. Movidius says that its technology can already spot if someone isn’t wearing a seatbelt or is talking on a cellphone. It can also identify the make, model, and color of a car. Together with unique features like stickers and dents, it can ID a particular car even if a license plate is missing or somehow obscured by matching it to another photo of a car where the license plate is visible.
A lot of people don’t like being watched, and they have ever more reasons to feel unsettled. This month, Georgetown Law School issued a study reporting that 117 million Americans, about half of all adults, are already logged in facial recognition databases used by law enforcement. AI security offers a modicum of privacy protection over other cameras, according to El-Ouazzane. “There is a lot of ability to obfuscate an ID element and just focus on the act which is actually taking place,” he says. That has the additional benefit of cutting down how much data needs to be transmitted and stored.
A smart camera would also know when to avert its eyes, he says. For example, a home security camera could be told to shut down in, shall we say, intimate situations. “I don’t want to give you too many examples, because as a French man I could get in trouble,” El-Ouazzane jokes.