Gait and Semantic Biometrics

Mark Nixon

(University of Southampton)

My talk will be phrased around the use of semantics in gait biometrics. As a biometric, gait concerns recognising people by the way they walk. As an application in computer vision, features are derived from sequences of images which can be used to recognise the walking subjects. I shall briefly describe the background to this work, the range of approaches and illustrate the progress in this new biometric.

Within this programme we have developing new approaches which use semantic concepts to reinforce conventional measures for pattern recognition. This has been deployed in biometrics, to improve recognition. Essentially, we augment biometric measures using measures derived largely by human intelligence. Using semantic descriptions is consistent with surveillance: for gait biometrics the semantic labels are those by which subjects are described orally. These concepts need to be labelled manually, so there are psychological considerations to be made, and there are many themes to the processing of this new data common with those in pattern recognition, including: feature set selection; classification; storage; fusion; potency; and efficacy.

We will show that by developing these semantic concepts, we can augment the feature vectors so as to improve recognition capability on standard datasets recognition capability by gait Further, we have shown how semantic concepts can be learned from the data, thus allowing object/ subject retrieval, as well as to enhance performance analysis. Further, this can relieve the manual labelling process. We show that humans can be retrieved from video recordings using descriptions notional characteristics, such as age and appearance, which is consistent with biometric search of surveillance material. We shall also describe how the
parameters can be learned from the data, so as to mediate the recognition process.

The talk is an extended version of the invited talk at IEEE BCC/ BiDS 2009
Tuesday 20th April 2010, 14:00
Robert Recorde Room
Department of Computer Science