Strategies for leveraging multiple footsteps in pressure-based gait verification

Abstract

The distribution of pressures under the feet during walking is unique to each person and can serve as a convenient biometric for identity verification at secured checkpoints. Although multiple footsteps are naturally captured as a person walks across pressure sensors, few studies have examined how this additional evidence can improve recognition robustness. This thesis evaluates several strategies for fusing information from consecutive footsteps under two constraints: (1) when only a single footstep can be assumed and (2) when at least two consecutive footsteps are available. Using six state-of-the-art deep learning architectures, fusion methods are explored at multiple stages of the recognition pipeline, including those that model stride-level relationships and asymmetries. Results show that while stride-level modeling offers modest benefits, simple fusion approaches provide substantial and consistent improvements across architectures, yielding more robust and generalizable systems for pressure-based gait verification.