DeepStep: A Deep Learning-Based Indoor Person Identification Framework Using Footstep-Induced Structural Vibration Signals

This paper introduces a large-scale, non-intrusive person identification framework using footstep-induced structural vibration signals. The increasing adoption of structural vibration analysis for person identification comes from its inherent non-intrusiveness and privacy preserving characteristics. However, existing methodologies are often constrained by the scarcity of extensive datasets, both in terms of the number of subjects and the temporal length of individual recordings, and frequently rely on supervised learning paradigms coupled with manual feature engineering. Consequently, the generalization capabilities of these approaches to broader populations are typically limited. To address these limitations, we have curated a comprehensive dataset of structural vibration signals acquired from 100 individuals. In addition, we have developed an unsupervised event detection method using features based on time, frequency, and wavelet analysis. Furthermore, we have developed DeepStep, a residual attention-based framework specifically designed for efficient feature extraction and classification of structural vibration signals. Experimental evaluation on our curated dataset demonstrates that the proposed approach achieves a Rank-1 accuracy of approximately 92% and a Rank-5 accuracy of approximately 96%.