Slashdot Mirror
In Istanbul, Cameras To Recognize 15,000 Faces/sec.
An anonymous reader writes "Istanbul's popular (and crowded) Istiklal shopping, cafe, and restaurant street is being outfitted with 64 wirelessly controlled, tamper-proof face-recognition cameras attached to a computer system capable of scanning 15,000 faces per second in a moving crowd for a positive match. The link from Samanyolu, badly translated by Google, states that 3 cameras are in place so far and that if trials are successful, this will mark the first time such a system, previously used by Scotland Yard and normally reserved for indoor security use, will be put to use in a public outdoor setting. It also notes that each camera controlled by the system is capable of 'locking onto' the faces of known criminals and pickpockets detected in the crowd and 'tracking' their movements for up to 300 meters before the next, closer placed camera takes over." Hit the link for more of this reader's background on the growing electronic encroachment on privacy in this city, which will be the European Capital of Culture in 2010, causing him to ask, "Is the historic city of Istanbul turning into the new London?"
While the article doesn't state it outright, it would appear likely that the outdoor face recognition system, if "successful," will be expanded to other crowded areas of Istanbul as well, which has already seen a dazzling increase in the number of installed plain-vanilla (non face-recognizing) CCTV cameras in recent years. This comes after Istanbul's two signature Bosphorus bridges have become passable only by vehicles with a mandatory vehicle windscreen-mounted electronic pass, subway and bus tickets in the city have gone electronic, vote tallying in municipal and national elections has become fully computerized, and future plans for mandatory biometric ID cards for all Turkish citizens have been announced by the government.
The ruling "moderate Islamist" AKP party appears to frame these and other e-government initiatives as "keeping step with the times," "keeping step with other major world cities," and "making living safer, easier and more efficient through the targeted use of electronic technology." Its secular critics, on the other hand, argue that everything and everyone under the sun is rapidly becoming electronically trackable thanks to the omnipresence of mobile phones and gratuitous overuse of these installed electronic systems, and that these systems will, eventually, form a dense surveillance grid that could turn daily life for Turks (and secular Turks critical of the current government in particular) into living in a veritable Big Brother House.
While the article doesn't state it outright, it would appear likely that the outdoor face recognition system, if "successful," will be expanded to other crowded areas of Istanbul as well, which has already seen a dazzling increase in the number of installed plain-vanilla (non face-recognizing) CCTV cameras in recent years. This comes after Istanbul's two signature Bosphorus bridges have become passable only by vehicles with a mandatory vehicle windscreen-mounted electronic pass, subway and bus tickets in the city have gone electronic, vote tallying in municipal and national elections has become fully computerized, and future plans for mandatory biometric ID cards for all Turkish citizens have been announced by the government.
The ruling "moderate Islamist" AKP party appears to frame these and other e-government initiatives as "keeping step with the times," "keeping step with other major world cities," and "making living safer, easier and more efficient through the targeted use of electronic technology." Its secular critics, on the other hand, argue that everything and everyone under the sun is rapidly becoming electronically trackable thanks to the omnipresence of mobile phones and gratuitous overuse of these installed electronic systems, and that these systems will, eventually, form a dense surveillance grid that could turn daily life for Turks (and secular Turks critical of the current government in particular) into living in a veritable Big Brother House.
In Istanbul but not Constantinople?
0x68ADA2CC
So I'm guessing that setting up a stand selling fake mustaches, Guy Fawkes masks, and Groucho Marx glasses on a busy corner in Consta... er.. Istanbul would get me a lot of money and a lot of police attention quickly.
Have you been touched by his noodly appendage?
CCTV ... ... ...
- less effective than promised
- doesn't reduce serious crimes like assault
- doesn't reduce, but shift crime scenes to other areas
- less effective than more light, more policemen,
- more expensive than more light, more policemen,
- often not working, tech staff admits
...64 wirelessly controlled, tamper-proof face-recognition cameras...
Sorry, but that's an oxymoron. It may be tamper-resistant (and some wireless devices have pretty good tamper resistance), but nothing that can be controlled wirelessly is tamper proof. Especially not when even the entity that has legitimate access (presumably the Turkish government) is entirely trustworthy to begin with.
To fight the war on terror, stop being afraid.
Sometimes technologies inevitably, if deployed, will be abused. Some technologies are too powerful to be in anybody's hands. Power corrupts, and all that.
...the phrase 15,000 faces per second is just an example of sensationalistic journalism.
There is a minimum input size for the identification of a 'face' dependent upon aspect and the focal length of the camera in question (amongst many other factors such as radial distortion due to the lens, whether the lens is shielded, lighting, et cetera); ergo, the camera in question, at a given focal length, could contain a field of view large enough and the resolution is high enough to meet 15,000 x the absolute mimimal pixel input for a 'face.' The processing for systems of this type (although I don't recall if it applies to this particular system) is tileable and accounts for boundary conditions (a 'face' falls on up to 4 tiles); therefore, the processing is highly parallel in nature. Most likely they meant that with the maximum cameras in place, given their proposed resolutions and fields of view, if they had unlimited computing power they'd theoretically be able to process 15,000 faces each second.
Solving a computer vision problem like this turns out to be highly hierarchical; i.e. a large number of computers process the incoming camera frames for optical flow, multigaussian motion detecxtion, edge detection, --insert motion map generating algorithm here--, these motion maps are shuttled to a second tier of systems who perform basic pattern recognition in order to discern probable aspect, reference points, and other forms of meta data. This tier can, if given a profile, apply discriminatory filters to sort the wehat from the 'chaff.' These 'probables' are then passed to the highest tier of systems who process this (hopefully) much smaller number of 'faces' using things such as color-space normalization from the original image, the motion map, and all the associated method data that has been generated along the way.
Luckily, most of the large companies working in these sorts of field are capable of producing crude prototypes; but, oddly enough, quality software engineers tend to be scarce amongst security companies. It is the startups and smaller companies (such as those found in Israel) that approach these types of problems with the flexibility to lead to some seriously scary big brother stuff.
Loading...
More importantly does it work on a moonlit night?
...to add to their massive data mining efforts. I can't even imagine the possibilities.
If they do, I hope for our sake it turns out better than their translations.
Somebody set us up the bomb.
Main screen turn on.
You have no chance to survive make your time.
For great justice.
Though those stories clearly refer to invading the privacy of one house, scholars universally extend to any prying.
I do not believe in karma. "Funny"=-6. Do good and forbid evil. Yours, Oft-Offtopic Flamebaiting Troll.
What's scary is that even with excellent success rates, that's going to be a lot of misfires. 15,000 faces/sec is 54 million faces an hour. At 'five nines' accuracy (which is far beyond what facial recognition can do as yet) that's still 540 false IDs per hour. It'd really suck to be one of those 540.
Rampant carbon sequestration destroyed the Dinosaurs' tropical paradise. I'm here to help repair the damage.
0.6% seems like a good ballpark figure for false positives.This research paper claims 0.6%. This article says "Commercial facial recognition technology ... had a 1 percent false positive rate."
15000 faces/sec * 0.6% false positives = 90 false positives per second.
How many cops does it take to ask 90 people per second to come to the police station to answer a few questions? How many busses does it take to take 90 people per second to the police station?
Once they get there, if it takes five minutes to look at each suspect's papers, run them through the computer, and clear them, that police station waiting room will need to be big enough to hold 27,000 people.
"How to Do Nothing," kids activities, back in print!
Anyone else think this is overkill? I can't pull 15,000 faces a second. Hell I don't think I know how to pull more than about 50 faces. Maybe 100 with variations. I can pull maybe 2 a second. Does this technology recognise middle fingers too?
These posts express my own personal views, not those of my employer
You're assuming every second that it rescans the crowd and does 15,000 new recognitions. More likely it scans the crowd constantly, and adds new faces to its database and continues to refine images on existing faces, tracking their movements to handle the interface between one camera and another.
it will be intressting to see if it really leads to an 1984 big brother state, or will actually lead to superiour crime fighting.
The problem is, most real crimes are rare and occur mostly in uninhabited or lower-class areas. This isn't going to stop murders, rape, major theft, etc. all the while eroding privacy. The summary mentions pickpocketing, pickpocketing is hardly observable in a crowd of people, I doubt these cameras would be able to track down the crime itself. Then there is the problem of false matches. A lot of people look remarkably alike in facial structure but look different in other areas that may or may not be tracked by this camera. For example, skin, eye, hair, etc. can often be the difference between a false match or a correct one. Machines though either rely on this too much (hair can easily be changed making it useless) or not enough (two people looking totally different with side-by-side comparisons but may have same facial structure), then some things can change in different lighting environments, etc.
Taxation is legalized theft, no more, no less.
...and this is sensationalism, there is no good way to process 15,000 faces/second.
... and those algorithms tend to get confused easily (the one on our test bench, from a very large company that does a lot of government sales, tries to chase shadows from the ceiling fan, inevitably follows them to the corner and never sees anything useful again until someone manually overrides it).
Here's how it works in the real world:
1) Face recognition demos well with small data sets. When you set it up on a conference room and scan everyone's face in the meeting, and then have each subject re-approach the camera, it works great. Note that the each subject in this demo is in the same lighting and didn't grow a mustache in the last 5min -- and there were only ~10 people in the data set. The real world is very different. A 15,000 subject data set is very very different.
2) When you set up a camera to scan for faces you need a lot of pixels in a head-on portrait type of shot. 640x480 is actually still pretty high resolution for a camera (there are some 8 megapixel ones but they are rare and they generate so much data that it quickly gets hard to switch and store that much data, even locally). Still, you'll need most of those 640 pixels wide in order to get a good shot of a face -- esp. if you're going to run that face against a large data set.
3) So, if you had 15,000 640x480 cameras, they'd still have to be setup in front of 15,000 turnstiles, or some other kind of crowd control device, for you to know that you're going to get a good face shot AND people would have to be moving through those turnstiles at 1 person/second. Picking faces out of a crowd? Not going to happen. You'd at least need PTZ (Pan/Tilt/Zoom cameras) with face finding/tracking/grabbing algorithm to even try
The closest centralized face-tracking technology is from a company called 3VR, they are used by banks to spot known bad-check writers at the bank counter (when someone cashes a bad check, they will use a different name/account/ID but they still show up with the same face). It works okay, better than nothing at least, but they can tolerate a lot of false-positives and just slow pay them or ask for a 2nd form of ID or whatever.
The company I started, Connexed, centralizes video from a lot of cameras, but I can say definitively that there is no tool on the market that will process 15,000 faces per second, no matter how much money you throw at it, and do anything useful other than trigger a flood of false-positive ID's faster than humans can process. You could always try to set the algorithm for maximum false-negatives (let a lot of bad guys get by) and minimum false-positives but even then, unless you have some way to get 15,000 people/second to look directly into a camera under good lighting, you're not going to have anything useful happen.
I'm sure it demo'd well, though, and the vendor got a good chunk of money for the trial that will ultimately fail.
My favorite quote doesn't fit into 120 characters. Now no one will like me.