Footstep pressure signal analysis for human identification

Pressure derived footstep signals are a growing field in biometrics, offering unobtrusive sample collection in comparison to established biometrics, with strong classification accuracy despite the highly variable nature of input instances. As a weak biometric, footsteps obtain lower predictive accuracy than stable alternatives and real world implementation will require reliable, yet flexible, feature sets that enable accurate class label partitioning. We suggest a method of retaining the spatial dimensions that are commonly lost during production of ground reaction force profiles and demonstrate the use of wavelet analysis on raw pressure signals for feature production. By analysing pressure signals obtained from common piezoelectric sensor arrays, we have trained a Random Forest classifier for individual prediction within a dataset of 10,413 footstep pair instances from 94 participants. Retaining spatial information for wavelet analysis returned error rates as low as 16.3%, showing strong predictive accuracy on a large, natural dataset.