This paper reports some novel experiments which assess the potential of footsteps as a biometric. We present a semi-automatic capture system and report results on a large database of footstep signals with independent development and evaluation datasets comprised of more than 3000 footsteps collected from 41 persons. An optimisation of geometric and holistic feature extraction approaches is reported. Following best practice we report some of the most statistically meaningful and best verification scores ever reported on footstep recognition. An equal error rate of 10% is obtained with holistic features classified with a support vector machine. As an added benefit of the work, the footstep database is freely available to the research community. Currently, the research focus is on features extraction on a new high spatial density footsteps database.
