Using AI To Stop Store Robberies In Real Time

While the world is busy helping retailers stop cybercrooks from stealing them blind online, every year 75,000 small retailers are being robbed at their physical locations.

Curtailing or even stopping it, Sean Huver, CEO and founder of Deep Science AI, told Karen Webster in the latest Topic TBD, is a matter of using the same visual search techniques that Google is using to enable a person snapping a picture of a cool blue bike to find out where to buy it.

Huver’s firm — “a humble startup” ready to open its closed beta to some fairly big names in the retail space — uses deep learning to identify what might be crimes in progress or on the cusp thereof, scanning and finding and pinpointing those who may be doing anything from concealing their faces to bringing out weapons to even menacing individuals at ATMs.

Deep Science AI has said that its platform allows for the monitoring of hundreds of security camera data feeds from retail stores and banks, in real time, alongside human analysis to identify a crime in progress, which alerts a human analyst to zero in on that feed and then dispatch the authorities in the hope of catching the criminal before they flee the premises.

This is possible, Huver said, because there is finally the match of technology advances that enables the serving of massive databases of images and finding and producing a match in real time. Over the last year, he said, much of the hardware that runs these sophisticated algorithms have become fast enough to enable real-time image searching.

And — now good enough to apply to physical crime against physical retailers — it begs the question posed by Webster: How do you build a data set that can separate the good guys from the bad guys?

“The first step,” said Huver, is “watching as many videos of robberies” that can be gotten hold of and analyzing what went wrong in each case. When they did that, what emerged, said Huver, was that about 85 percent of the thousands of robberies that happen every year — or more than 200 daily, given an annual 75,000 pace — take place at smaller retail businesses (a data point that sits alongside the several thousand bank robberies that still occur annually on premises). Many of these crimes at smaller establishments — 70 percent of them — remain unsolved, because the call for help comes after the robber has left the premises.

Though Huver stated that high-definition surveillance cameras are pretty much everywhere, most of them don’t actually help the retailer the moment a crime is committed — or remain useful for evidence after the fact.

Conversely, deep learning and Deep Science AI’s technology can pinpoint certain events and objects, such as a weapon being produced in front of a cash register or someone covering their face as they make their way through a store (as they are mindful of cameras).

And, he noted, the platform continuously gets smarter as it gets more data in hand, applying that data to real world solutions in real time.

With that real-time data, said Huver, when one of those detections is made, a monitoring analyst, who is remote at a partner company, is alerted in real time. The determination is made as to whether there is a real crime in process or whether a false positive has been triggered. The platform itself learns how to recognize false positives, said Huver, amid the interaction with human monitors, who make the case-by-case determination (anecdotally, retail scanning guns, said Huver, have driven false positives in the past, with obvious alerts that real firearms are being brandished).

At the ATM, where Webster noted not all crimes take place with a gun, Huver said his company can identify when there is fighting or people struggling over money, for example, with a view of identifying behavior, in addition to objects signaling that something is wrong.

Huver said that much of the work in AI has to be done with forethought and foresight, in terms of knowing where there are images to be gleaned and how information can be added to a system once new data arrives — in short, how the system can learn even as it analyzes.

The value proposition, continued Huver, is one that centers on faster response times. The average robbery at a retail business in an urban area, he said, lasts around two minutes. But there are cases where it goes on longer, say, where a robber locks an employee in a bathroom or where they seek to gather up as much in ill-gotten cash and merchandise gains as possible. The upshot in this case is that the police can get there much faster or stop a robber in his or her tracks.

Another part of the value proposition, said Huver, “is that we believe this is going to be a great new deterrent.” The surveillance camera is destined to become ubiquitous, said the executive, and even though the deterrent effect is not absolute, there is indeed some impact if there is awareness of monitoring. “We believe that as this technology becomes more widely known about, people will be less likely” to commit crimes.

A third value proposition, said Huver, ties in with liability, where “if you are a business owner and you’ve been robbed before and you do not take additional steps” to improve security, and a customer is injured in a subsequent robbery, liability is in place … which may indeed spur more retailers to beef up their security efforts.

Noting that 30 percent “clearance rate,” where crimes are solved and criminals apprehended, Huver stated that a goal in place to get that rate up to 40 percent or 50 percent means that “deciding to rob a business becomes a much scarier proposition.”