Sunday, April 22, 2007

Biometrics: Gait recognition accuracy rates in controlled conditions

Researchers at the University of Southampton (UK) reported a 100% accuracy rate for verifying the identity of individuals by automatically analyzing the way they walk. IT Week (a UK information technology magazine) has some details.

Based on the IT Week article, it appeared to me that all the tests were conducted under controlled conditions. Visually-based biometric techniques generally require good lighting conditions (i.e. proper exposure of the image) and resolution. They also require that the features needed for recognition are not occluded.

Why would one need gait recognition when face recognition already exists? Some obvious advantages are that gait recognition doesn't require that the face is oriented toward the camera and the image resolution requirements are not as high. But if the conditions are controlled anyway, why can't the person be required to be close to and facing the camera?

A couple of questions that come to mind are 1) how much less the spatial resolution can be and 2) how robust it is to image compression (e.g. MPEG-4)? These could give the technique additional advantages.

One final observation I should make is that, as with all biometric techniques, being able to fuse the results of more than one technique improves the system accuracy and robustness. That alone could justify the inclusion of the technique in fielded systems.

3 comments:

Anonymous said...

Interesting.

Here's their IEEE paper from the end of last year.

http://eprints.ecs.soton.ac.uk/13369/01/nixon_pieee.pdf

Here they don't claim 100%. I haven't read their paper yet, but glancing through it, their ROC curves are impressive. The 100% might just be an over-eager journo looking at only false accepts or false rejects.

Still, very interesting. Wonder what 'within-person variability' they consider.

-a.a

Anonymous said...

"One final observation I should make is that, as with all biometric techniques, being able to fuse the results of more than one technique improves the system accuracy and robustness"


If you use two or more independent biometric technologies in the same system or device, e.g. iris scans and facial recognition and fingerprints, as has been suggested for the UK National ID Card for instance, then surely the overall False Positive rate for the whole system is made worse not better than the False Positive Rate of each component used individually ?

Multiple biometric techniques allow for a wider coverage of the general population, since there is no single biometric technique which covers 100% of the population (there are people without one or more eyes or fingers etc.).

However a reduction in the number of "insufficient data" runs, does not necessarily mean that the overall system accuracy is better.

Keith said...

Dear Watching,
Good point. What you say is true IF one uses OR logic (either biometric method registering a 'positive' means that the combined result is 'positive') in the system design instead of AND logic (both must register 'positive' for the combined result to be 'positive').

Of course, the OR case is a simple one and maybe, when put into practice, intended to maximize the number of positives, false or not, under the idea of impressing people (lies, damned lies, and statistics) or under the assumption that it is better to 'flag' anyone remotely suspicious and subject him to additional security screening or other measures, as the application dictates.

Of course, one could use some more sophisticated means of fusing the data that takes into account how 'confident' one is in the different metrics, what the environmental situation is, what the desired false negative/positive threshold is, etc. This could substantially reduce the number of false positives.

As you can tell, I am approaching this from the standpoint of systems engineering - i.e. how to design a system to meet the desired system specifications. What comes before (how those specifications are written) and after (how the system is fielded and how the results are used) are subject to policies and procedures. All three aspects (policies, procedures, and system design) come into play in this very complicated area, not to mention all the other secondary factors that affect them.

Regards,
Keith