Spatial footstep recognition by convolutional neural networks for biometric applications

This IEEE SENSORS 2016 paper trains a convolutional neural network to recognize people from spatial footstep patterns recorded by a floor sensor. It was tested on a large footstep database from over a hundred people.

Presented at IEEE SENSORS 2016 in Orlando, Florida, by Omar Costilla-Reyes, Rubén Vera-Rodríguez, Patricia J. Scully, and Krikor B. Ozanyan, the paper proposes a convolutional neural network that learns spatial footstep features end to end from a floor sensor system for biometric applications. A biometric system identifies or verifies a person from a measurable body trait, here the pressure pattern of a footstep. The paper was the earliest of several collaborations between the University of Manchester and Vera-Rodríguez’s biometrics group at the Universidad Autónoma de Madrid, using the SFootBD footstep database.

Generalization was assessed with independent validation and evaluation datasets drawn from that database, which the authors describe as the largest footstep database to date, with many thousands of signals from over a hundred people. The network preserves the spatial structure of each footstep measurement to learn features associated with a person’s walking pattern.

The paper reports footstep recognition performance as equal error rate (EER), the operating point at which a system wrongly accepts and wrongly rejects people equally often, so lower is better. The reported error rates are low and improve on previously reported footstep recognition in the spatial domain, and the paper examines how data partitioning shapes biometric performance. Its benchmarks provide a baseline for comparing spatial footstep-recognition methods. The follow-up 2018 journal paper extended the approach to combined spatio-temporal representations with deep residual networks.

Dr. Costilla Reyes presented the work as a poster at the conference, inviting attendees to learn how footsteps can serve as a security verification signal. The paper forms the fifth manuscript of his PhD thesis.