L.E. agencies fine-tune face ID
By THOMAS FRANK
In states such as Florida and Virginia, law enforcement officers experiment with face-recognition technology and uncover systems' pros and cons.
Police in one of Florida's largest counties are hoping that a digital camera may be the crime-fighting weapon of the future.
The $200 cameras do more than photograph people questioned on Pinellas County streets along Florida's west coast. Deputy sheriffs transmit a mug shot from car laptops to a database holding 4.5 million digital photos of people who've been arrested in Pinellas and five other Florida counties.
Since 2004, Pinellas deputies have identified thousands of people who initially gave them false names. And they've arrested 254 people found to have warrants.
"They wouldn't have been arrested if we didn't have the system," Pinellas Lt. Jim Main said.
Pinellas is pushing the frontier of face-recognition systems, an innovative and potentially privacy-invading technology that converts facial images into digital data that can be stored and analyzed in high-speed computers.
The technology has been hyped since 9/11 for its potential to instantly spot terrorists in airports or to identify criminal suspects captured on surveillance cameras.
Recent tests, however, have shown the technology to be less RoboCop and more the Invisible Man, and major shortcomings in surveillance have left authorities scrambling. Technology limitations have left police with images that computers can't always read accurately and might cause the computer to misidentify an innocent person as a terrorism suspect.
"The technology created a lot of false expectations that still exceed its ability to perform in a 'hostile' environment where you can't control the lighting or someone's pose," said Walter Hamilton, chairman of the International Biometric Industry Association, a trade group.
Virginia Beach police discovered the technology's flaws shortly after they began using it in 2003 in a test funded by the Justice Department.
Three cameras positioned above the main commercial strip of Virginia's largest city compared pedestrians' faces with facial images in a photo database of 1,000 local fugitives. Comparisons took a quarter-second. An alarm sounded in a video-monitor room at Virginia Beach police headquarters when the computer detected a match.
If an officer confirmed a match by comparing photos of the pedestrian and the fugitive, another police officer would be dispatched to Atlantic Avenue, where the pedestrian was photographed.
To the department's surprise, the system began sounding up to 300 alarms a night. To its annoyance, all of the supposed matches were quickly found to be no match at all, said Dave Kent, Virginia Beach master police officer.
"It was a distraction," he said.
The system produced so many false alarms because its "match parameters" were set at a low level to compensate for flaws in the photos it took, Kent said. Sometimes the system sounded an alarm when it spotted a T-shirt decorated with a face, he said.
After three years refining the technology, Virginia Beach shut the system down in 2006. "There was not a single arrest," Kent said.
The technology is even more unlikely to spot terrorists because the database photos of them are often old, said Takeo Kanade, director of Carnegie Mellon University's Robotics Institute. "You not only have facial changes over time, but changes in hair pattern will affect recognition," Kanade said.
The technology functions by measuring distances between about 100 points on a face. Those distances are converted into an algorithm, a unique sequence of thousands of 0s and 1s that are stored in a database.
When a Pinellas deputy sends a photo to the county system, he gets back an e-mail with about 20 photos of possible matches, said Scott McCallum, a system analyst in the sheriff's office. One of the photos is usually the obvious match, he said.
Pinellas deputies get good photos because their system is "cooperative" -- they photograph only people on the street who give their permission. Photos also are taken of suspects as they are booked.
"This works because we can control the environment," said Dan Wiggins, support-services director for the sheriff's office. The system "helps tremendously" in preventing suspects from giving someone else's name when they are booked.
The Illinois secretary of State's face-recognition system has stopped more than 3,000 people since 2000 from using false names to get driver's licenses or state ID cards, said Dennis Krier, training coordinator for drivers' services.
The state photographs license applicants and instantly compares the photo with 20 million images of people who have licenses or ID cards.
"It's definitely been a deterrent to identity theft," Krier said.
Face-recognition technology has improved dramatically in recent years, according to a National Institute of Standards and Technology study released in March. The study found technology tested in 2006 had an error rate of just 1% to 2%, down from 20% in 2002, study director Jonathon Phillips said.
For people photographed in poor conditions -- in hallways and outdoors -- the 2006 failure rates ranged from 13% to 18%.
"The angle of the face, lighting, facial expressions and disguises are all issues for surveillance," said Victor Lee, a consultant at the International Biometric Group.
L-1 Identity Solutions of Stamford, Conn., got $1.8 million in December from the Homeland Security Department to improve the technology's surveillance abilities. The company is trying to build software that creates a picture of a complete face from digital photos that show only part of someone's face. Partial facial images often show up on surveillance cameras.
"If you could construct that face, you increase dramatically the likelihood that you could identify the perpetrator," said David Boyd of the department's Science and Technology branch.
The technology also could allow innocent people to be identified by cameras, said Melissa Ngo of the Electronic Privacy Information Center. "That raises a lot of questions about your ability to be anonymous in public," she said.
Copyright 2007 USA TODAY
L.E. agencies fine-tune face ID