Slashdot Mirror
Video Surveillance System That Reasons Like a Human
An anonymous reader writes "BRS Labs has created a technology it calls Behavioral Analytics which uses cognitive reasoning, much like the human brain, to process visual data and to identify criminal and terroristic activities. Built on a framework of cognitive learning engines and computer vision, AISight, provides an automated and scalable surveillance solution that analyzes behavioral patterns, activities and scene content without the need for human training, setup, or programming."
Nothing can go wrong!
It's a press release pretending to be journalism.
If it doesn't need training, how does it define "terroristic activity"? Is it the "I'll know it when I see it" definition?
The article seems to indicate it works like a Bayesian filter on the video - pointing out things that aren't typical for the camera.
Much like any automated system that is supposed to filter out false positives, it is probably pretty easy to train either the operators or the system itself to throttle back the sensitivity to a point where it ignores everything.
The "machine learning engine" is a "datacenter" (warehouse) full of cheap African laborers who are all watching the cameras.
(this is a joke, it just isn't funny, and it is meant to illustrate a point. See the next line):
God/nature/FSM/evolution/al gore/$deity has done a pretty damn good job at building our brains, why are we trying to reinvent that wheel in a computer?
NewslilySocial News. No lolcats allowed.
My guess is it applies a few simple heuristics to analyze the behavior and the real trick is identifying the behavior.
Example: In an alley behind a hotel people frequently walk out a door, put something in a container, and walk back in. This becomes "normal". Then someone goes out back and starts smoking. Whoops, wtf is this! Alert, alert. OK, so this gets flagged as OK a few times. The system decides it's OK. However, when two people hold a third at gunpoint and linger in an area of the alley not usually used for smoking, this would now trigger as abnormal.
Another thing it might notice is the same person coming back to the front of a convenience store, waiting a minute, then leaving, then coming back again. Most people only walk in, walk out - this is abnormal.
So it won't tell you someone is burglarizing you, but it might focus your attention on a camera where something could be happening. I'd assume it would get better over time as things were flagged "ok" or "not ok", but at best it would provide some simple pre-filtering to focus human attention on scenes that are slightly more likely to be "interesting".
What a great way to absolve any personal responsibility. Detained wrongfully? Not our fault, the machine said you were moving like a terrorist.
Much like detecting terrorists by facial recognition, this is vaporware until they publish some numbers.
I once had someone misplace a sales call to me, being proud his facial recognition system was 70% accurate. He had no idea how much his system is a pain in the ass when its wrong, and for the airport security business he was trying to get, 90% accuracy is considered terrible.
Mod parent up. Said AI first needs to distinguish between "activity" and "the wind blew a leaf across the screen". Then you need to distinguish between "lights a cigarette" and "lights the fuse on dynamite".
So, if it already does all that, just one more question: how do you define "criminal and terrorist activities" programmatically when not even the law is clear? Even shooting people can be a non-criminal act.
It must first differentiate between "time flies like an arrow" and "fruit flies like a banana". Then, and only then, can be the system be trusted.