Sunday, May 9, 2010

Automated Video Image Analysis Will Be the Next Big Breakthrough

Remember Chance the Gardener character in Being There? "I like to watch TV" was his famous line. The problem with most business videos is that we produce a lot of unwatchable footage, especially those taken by security camera systems. And there aren't any Chances around that want to watch them, either.

The trouble is that trying to find the one or two actionable events in all of that footage isn't easy. As an example, take the story in today's NY Times that mentions how the military is being buried under a massive pile of video footage from its Predator drones that are flying over Pakistan and Afghanistan. The situation is only going to get worse, as newer drone models will be sending streams from dozens of cameras within a few years.

The story has already been told about how the drones' video feeds are available to anyone who has some minor software skills - the actual control channels are encrypted but the video transmissions aren't. This is because many legit people need to see what they are broadcasting and the military hasn't been able to implement any encrypted viewing packages on these streams.

In the Times article, a bunch of soldiers based in Hampton, Virgina sit in front of the screens and see it in real time, and then make screengrabs available to the right people via computer chat rooms. I hope for the sake of everyone involved that these chat rooms are encrypted, but the article didn't say.

So how do we implement our automated Chance Gardner? There are a couple of technologies that can help here, but they aren't easy or cheap to implement. One is the telestrator, the device made popular by John Madden and Monday Night Football where a commentator draws on the screen and you seen colored squiggles to highlight what is going on. The ones that Madden and the pros use are very expensive, but there are dozens of products available for the PC market, with some freeware products such as VideoMage Producer.

The telestrators are nice, but again, someone has to be watching the video and doing the electronic doodling. You need more than the fast-forward button to do this - ideally, you want some kind of automated system that can identify actionable moments on the video. This is what the next class of products does, called intelligent image analysis. They have computers to look at the stream and highlight particular activities that a human operator can come back to and review later.

This is what the company is doing with its retail checkout analysis systems. Typically, a retail store installs video cameras above each checkout aisle and records what the checker is doing as items pass through the point of sale scanning devices. There are all sorts of scams that can be used, such as "sweethearting" (a confederate is giving free items that aren't scanned) and looking like you are scanning a bar code when you are just passing the item around and over to the bagger. So what is needed is a system that ties into your point of sale and can flag when these items aren't rung up at the register. I got to see a demo last week and thought this was way cool. The company claims their software can have a six-month ROI and significantly reduce the cost of stolen goods. And the good news is that no one has to watch all the security tapes to see those few sweetheart moments.