The security market has seen its share of technologies that scream “My time has finally arrived!” For facial recognition technology, most of us are still waiting; however, advances in accuracy and the fact that it has found applications in government and municipal law enforcement settings has me thinking that the technology’s “time” is about to arrive.
Facial recognition (F.R.) systems are computer-based security systems that use “geometric mapping” and other techniques to analyze and categorize human faces by distance between the eyes, shape of cheekbones, etc. These measurements are stored in a database or used to compare with profiles stored for establishing a match.
The technology sounds terrific for security; however, its time has not come because of limited effectiveness and accuracy. Issues include masking features, such as glasses, facial hair or expression; poor lighting; and processing power for on-the-fly analysis. I participated in a market survey of integrators and consultants last year — no surprise when all the respondents said they were neither using nor specifying the technology.
Still, thanks to noticeable improvement is the accuracy of F.R. systems, the technology is inching toward prime time. One major enhancement was the conversion of 2D images to 3D — thanks to companies such as CyberExtruder (www.CyberExtruder.com), which can convert a few 2D images of a person’s head/face into a unique 3D model. The 3D model can then be rotated to get an improved frontal view to present to the image processor.
Algorithm developers, such as NEC (www.necam.com/Biometrics), have attacked other weaknesses of the technology, such as masking features and expression, and have pursued non-geometric techniques. An example of this is pattern analysis — a technique that quantifies skin texture and its unique lines, patterns and spots.
Advancement in wide dynamic range cameras has reduced the effects of lighting variation, and megapixel cameras provide increasing richness of detail.
Periodic studies by the National Institute of Science and Technology (NIST) show a remarkable progression in accuracy. In published studies, the false non-match rate (FNMR) fell from 79% in 1993 to 0.3% in 2010, achieved by NEC.
This technological advancement has led to the use of face matching as a forensic and authentication tool for law enforcement. Today, the U.S. State Department operates one of the largest F.R. systems in the world — with more than 75 million photographs used forvisa processing. The German Federal Police employs F.R. on a voluntary basis, to enable people to pass through the completely automated border security system in the Frankfurt Airport. The technology is also extending beyond law enforcement. In Ontario, Canada, casinos are deploying F.R. after testing demonstrated a 91 percent success rate in identifying problematic gamblers.
Still, the technology is not a fit for every application, especially in cases where a number of large images need to captured, processed and analyzed in real-time for response and intervention.
With large camera counts operating at multiple frame rates per second — perhaps with multiple faces per frame — real-time processing and network bandwidth needs can be overwhelming. For example, with images from 100 cameras producing one face per second and matching a database of 100,000 faces, 10 million comparisons per second are required — still overwhelming for even the most complex computer.
But the industry is working on that issue as well. Companies like FaceFirst (www.FaceFirst.com) are addressing this through a variety of techniques. First, analytics on the edge can be used to perform pre-filtering to select the best facial rendering, reducing both network load and head-end processing. By employing a “pipeline” of algorithms at the head-end and paying attention to human usability, the analytics aim to present the operator with faces demonstrating the highest-probability match.