Contrastive Sensor Excitation for Generalizable Cross-Person Activity Recognition

Unsupervised Domain Adaptation (UDA) has been proven effective in extracting domain-invariant features, facilitating successful model generalization from the training user to the new user in sensor-based Human Activity Recognition (HAR). However, such extraction of the domain-invariant representation that only focuses on aligning global features cannot be robustly applied to Multi-sensor-based HAR (MHAR) due to the diverse distributions inherent in different sensors. In this paper, we propose CONtrastive SENsor eaxciTation (CONSENT), a novel unsupervised domain adaptation framework for cross-person MHAR. CONSENT considers both sensor-specific feature alignment and global feature alignment. At the sensor-specific feature level, we propose an effective contrastive loss to ensure consistent distributions among identical sensors across domains without applying any pre-defined augmentations. Meanwhile, to further minimize domain discrepancy at the global feature level, we design a simple but effective excitation mechanism, which motivates the model to pay more attention to sensors that are critical for activity recognition. In this way, the model is encouraged to learn representations that are robust to sensor-specific information. Empirical results demonstrate that CONSENT achieves state-of-the-art performance on three public MHAR datasets.