South Korean Study Delivers a Face Rec First (90%+ Accuracy Using Surveillance Video)

Those following the South Korean government’s most recent biometric initiative know it to be one of the most ambitious facial recognition testing and deployment projects ever conceived. The project, in its various phases, has been featured quite extensively on Korean television here and here.

Under the project, the Korean National Police (NPA) sought to deploy facial recognition technology against Seoul’s plentiful video surveillance cameras, both to generate alerts to the presence of targeted individuals, as well as provide a mechanism to quickly search city surveillance archives for persons identified after the fact.

Though the NPA has been quite public on its facial recognition project in general, details regarding the specific technologies and testing results have been closely guarded. However, today after much anticipation, the NPA has finally made its initial findings public. According to the NPA and SK Networks spokesmen, 3VR's facial recognition platform demonstrated the highest percentage of accuracy of any of the eight tested technologies in trials run by the NPA and its testing partner, South Korea's IT giant SK Networks.

Indeed the results, at around 90% accuracy, would seem far better than those achieved in any previous public video facial recognition study. After a grueling multi-year testing process, in 3VR SmartRecorders and SmartCams provided between 85 percent and 92 percent accuracy in recognizing and matching faces in a few crowded, highly-trafficked public train stations in Seoul. In each case, the images analyzed were of fast-moving groups of commuters entering or exiting various transit areas en masse.

Said Sung-Ho Kong of SK Networks:

“In 2008, we performed a live, uncontrolled test of 3VR's facial recognition technology in Seoul subway stations, where the solution was an impressive 85-92 percent accurate, depending on conditions. No other solution approached this level of accuracy, vastly improving our ability to track, find and thwart crime in subways and other highly populated areas, which had previously proven extremely difficult to monitor.”

To better appreciate just how impressive the NPA’s results are, we need to compare them to the best previous facial recognition study. In 2006-2007, a similar facial recognition-use case was evaluated by the German Federal Police. After months of testing in a German subway with lighting and traffic much more controlled than in South Korea, a mediocre 60 percent accuracy rating was achieved.

The improvements in accuracy afforded by 3VR are the result of a unique approach to video facial recognition. While most face rec vendors offer technologies optimized for the comparison of flat, normally lit, passport-style photos, only 3VR's software was built from the ground up to address many of the unique challenges presented by real-time surveillance video.

According to Tim Frederick, director of engineering at 3VR:

“This study demonstrates the breakthrough power of 3VR’s patented facial surveillance technology. Unlike other attempts at high-volume face surveillance, which re-purposed still-image face recognition algorithms, the South Korean study benefited from 3VR’s end-to-end video analysis system, specifically designed for this type of demanding video application.”

Expect more details from 3VR and the NPA shortly, but in the meantime, please check out SDN's Rhianna Daniels' feature story on the announcement, as well as a quick diagram of how the technology works and even some up-close and personal screen grabs of the UI.

UPDATE June 22, 2009:

Here is a translation of the South Korean National Police's Phase 1 testing results: