An experimental study on the feasibility of footsteps as a biometric

This paper reports some experiments concerned with footsteps as a biometric. We present a comparison between different feature extraction techniques as well as classification methods, obtaining the best results with holistic feature extraction and a support vector machine (SVM) classifier. Results are reported in terms of detection error trade-off (DET) curves, showing minimum equal error rates of around 10%. Previous experimental work has limitation in sizes of databases; here we have over 3000 examples across 41 persons and are therefore able to design independent development and evaluation datasets. Once finished and validated, this database will be made freely available to the research community.