Open Access Open Access  Restricted Access Subscription or Fee Access

Physics-Guided Model Transfer for Human Gait Monitoring using Footstep-Induced Floor Vibration

MOSTAFA MIRSHEKARI, JONATHON FAGERT, SHIJIA PAN, PEI ZHANG, HAE YOUNG NOH

Abstract


In this paper, we present a physics-guided model transfer approach for robust human gait monitoring across various structures using footstep-induced structural vibrations. Understanding and characterizing the structural vibration caused by human footsteps provides insight about their gait, which is important in smart healthcare applications. Compared to current sensing approaches for gait monitoring (e.g., pressure-based, and mobile-based, and vision-based sensing), vibration sensing is sparse, non-intrusive, and causes less privacy concerns. Conventionally, finding gait-related information using structural vibrations occurs through data-driven inverse modeling with supervised learning, which extracts information about human gaits based on the vibration measurements and corresponding labels. However, the measurements are also affected by the structure properties; therefore, the models developed in one specific structure for mapping the measurements to labels are not applicable to other structures. This results in extensive calibration and training requirements. To overcome these challenges, we present a physicsguided model transfer approach that projects the vibration data into a space where the effect of structural properties is minimized. To this end, we first analytically show that the structural effect in the signal is correlated with the Maximum-Mean-Discrepancy (MMD) data distribution distance across various structures and then, minimize the MMD to find a projected space with reduced structural effect. This reduced structural effect enables us to train a footstep model in one structure (where labeled data is available) and use this model for labeling the samples in the other structures with no labeled data. In field experiments across two types of structures, our approach has achieved 0.95 and 0.97 F1-scores for footstep detection, which is up to a 9X improvement when compared to the baseline approaches.


DOI
10.12783/shm2019/32337

Full Text:

PDF