Gait recognition made possible for very low-resolution video
Wed 15 Jun 2016
A group of Spanish researchers have used the power of neural networks to create an effective system of fingerprinting the way that people walk – even at resolutions as astonishingly low as 80 x 60 pixels.
The emerging biometric technology of gait recognition is among the most controversial methods of ‘fingerprinting’ individuals – ironically because it is among the least invasive, most passive and most undetectable techniques. Researchers from Munich hit the headlines in 2012 with a gait recognition system that employed Microsoft’s Kinect sensor, whilst Chinese scientists have developed an app that claims to individuate a person’s mood by their gait.
However the Spanish paper [PDF], featuring researchers from three universities, outlines an effective technique using convolutional neural networks which not only takes on the challenging TUM Gait from Audio, Image and Depth (GAID) database, but works as well when the video in the database is reduced to one-eighth of its usual resolution.
Previous work, claim the researchers, has required hand-crafted features within crafted systems to achieve optimal gait recognition, but the CNN model that they have developed works without bespoke attention, extracting high-level human features that can be distilled down to a ‘gait formula’, or individual fingerprint.
The initial pass with the system achieved a recognition accuracy of up to 99.6%, with an average of 98%, regardless of differing clothing or shoes worn by the subjects in the database. However the second of the two result-sets required the system to re-identify subjects after a period of some months had elapsed – a general challenge in the field of gait recognition – and here accuracy topped out at 89.6%.
The paper also notes that the accuracy with which the system can identify females vs. males is considerably lower (77% vs 96%), though this may be due to the peculiarities of the GAID dataset.
Nonetheless, these are remarkable outcomes considering the extent to which the researchers were determined to handicap themselves:
‘Comparing our best results with previously published ones, we observe in Tab. 3 that our accuracy (rank-1) is on a par with those methods, even though we are using video frames with a resolution eight times lower than the others.’