Shadow Analysis Could Spot Terrorists 245
Hugh Pickens writes "An engineer at Jet Propulsion Labs says it should be possible to identify people from the way they walk — a technique called gait analysis, whose power lies in the fact that a person's walking style is very hard to disguise. Adrian Stoica has written software that recognizes human movement in aerial and satellite video footage by isolating moving shadows and using data on the time of day and the camera angle to correct shadows that are elongated or foreshortened. In tests on footage shot from the sixth floor of a building, Stoica says his software was indeed able to extract useful gait data. Extending the idea to satellites could prove trickier, though. Space imaging expert Bhupendra Jasani at King's College London says geostationary satellites simply don't have the resolution to provide useful detail. 'I find it hard to believe they could apply this technique from space,' says Jasani." Comments on the article speculate on the maximum resolution possible from KH-11 and KH-12 spy satellites.
Re:Upon deployment.... (Score:5, Informative)
Unmanned Aerial Vehicles instead of satellites (Score:2, Informative)
Defeating gait analysis (Score:5, Informative)
Re:Upon deployment.... (Score:5, Informative)
Aye, well the Scots have them all beat:
University of West Scotland research reveals that a woman's gait may reveal her orgasmic ability. [physorg.com] - A new study found that trained sexologists could infer a woman's history of vaginal orgasm by observing the way she walks. The study is published in the September 2008 issue of The Journal of Sexual Medicine, the official journal of the International Society for Sexual Medicine and the International Society for the Study of Women's Sexual Health.
Combine that with satellite-based shadow analysis, and... Giggity!
Overblown story (Score:4, Informative)
Making the training data useful for recognition is challenging enough. If you have footage of a person walking against a white wall at a controlled distance, it is easy to gather this data. However, if your training data is from a video of a person walking through city streets, much less a market place, there is an awful lot of human processing that needs to be done in order for the data to be useful for training. Also, as with facial recognition and other visual recognition techniques, gait-recognition is highly susceptible to changes in camera angle. If you train a gait-recognition algorithm on images of someone walking towards the camera, that doesn't mean that you can identify them with any reasonable success from the side or above. In essence, in order for this to work, you would need an ample amount of training data on a terrorist in a controlled environment. That probably isn't very likely. As you'll notice in from the article, these experiments were conducted in a specific controlled environment.
This story strikes me as someone doing some interesting research, but I'd be curious if we get any meaningful results from this work, even 10 years down the road.
Well, I am a computer vision scientist... (Score:4, Informative)
Well, IIACVS (a computer vision scientist), so here goes...
No, this won't spot terrorists. Currently biometrics are pretty good at answering the question "is this person Mr. A?" where A is know, and providing a yes/no answer.
They're OK at answering the question "Which of A,B,C,D is this person?", up to a fair number of people.
What they suck at is "I A any of the potentially 6 billion people who might go past this camera?"
The reason they're good at the first, is becaus eif you want to get entry based on biometrics, you don't generally hide your appearance. In the last case, hiding appearance s easy. Basically if you wear a large sack and sunglasses and gloves, your face, irises, fingerprints and gait are not accessible. No fancy camerawork will help with that and the system will not work.
Then there's more minor things like beards/lack of for faces, and for gait, a stone in the shoe, leg injury, John Cleese, an embarrassingly placed itch, and so on which also throw off the system.
Basically, Humans have had millions of years to perfect (and a large chunk of brain dedicated to the task of) identifying people we know. We're eally good at it, and can identify people we know very easily in a reasonable sized group.
We still don't scale well up to very lage groups, probably because the problem is too ill poosed to be tractable.
Re:Geostationary? (Score:4, Informative)
Apparently they might well be able to, and may already have done so:
http://physicsworld.com/cws/article/print/19415
Re:Yes, but that still doesn't answer his question (Score:4, Informative)
Your gait isn't so much like a fingerprint.
A little background. I am not familiar with this research specifically, but I have exposure to a very similar concept.
I used to do ground search and rescue in New Jersey. A big portion of what we did was woodland SAR for missing persons. Because of this, there was a lot of emphasis put on tracking.
There is a man named Tom Brown Jr. who is basically considered the modern foremost expert on the subject. He learned the basics from a young age from an old Apache scout who was his best friend's grandfather. It sounds incredible, but the man has written several books, both technical, and biographical. The technical ones aren't of much interest to folks who don't have an interest in tracking, but the biographical ones I would highly recommend. He currently runs a school in Toms River, NJ [trackerschool.com].
The organization I did SAR with put a lot of stock on Tom Brown's methods and incorporated them into their training schools. Eventually we opened up a dedicated tracker school, though I never participated in that level.
There is a technique known as pressure release tracking, where one looks at the characteristics of a track in a soft medium like sand, mud, or to a lesser extent, gravel or such. Within the track exists a whole environment that was created by the state of the organism that made it. Most people can figure out that if you shift your weight to your left, or favor your right foot, or are limping, that you'll see that in a track. But you can also see other things. Is the subject hurt? Is it hurt somewhere other than the legs? Is it tired? Is it male or female? Is it pregnant? How much does it weight? How tall is it? Is it carrying something? Does it have to urinate? Is it sexually aroused?
I know people who have reached the level where they can infer these things accurately. To me, it's not a stretch to believe that there are other ways that this could be done (this shadow technique for instance).
A good tracker can tell a lot by looking at your tracks, so I'd so I don't know how they plan to use gait data in a useful way, but I'm willing to entertain the idea.