Few-Shot Learning for Cross-Domain Human Activity Recognition Using Wearable Sensors.

Abstract

Sensor-based human activity recognition (HAR) plays a critical role in pervasive health monitoring, smart home, and sports tracking. While deep learning excels in HAR tasks with abundant labeled data, its real-world deployment is hindered by the scarcity of annotated data and significant domain shift arising from variability in individuals, sensor positions, and scenarios. To overcome the aforementioned challenges, a novel lightweight cross-domain few-shot (FS) sensor-based HAR network (CFSH-Net) is proposed for cross-domain activity recognition with limited labeled samples. CFSH-Net leverages wavelet-driven depth-wise bilateral decomposition (WDBD) as the feature extractor, significantly enhancing the discriminative ability of the FS classifier while maintaining low complexity. Crucially for personalized applications, CFSH-Net adapts to unseen activities and domain shifts across sensor positions, users, and datasets using only a few labeled target samples and no fine-tuning process. Extensive experiments on six public benchmarks, comprising OPPORTUNITY, PAMAP2, UCI-HAR, WISDM, WISDM2019, and USC-HAD, validate its effectiveness. Under the 5-way 5-shot setting, CFSH-Net obtains 85.22% and 88.66% accuracy for unseen activities of PAMAP2 and OPPORTUNITY, respectively, and reaches 79.94% average accuracy in cross-position recognition. It further demonstrates strong cross-user generalization on PAMAP2 and USC-HAD, and stable cross-dataset transfer when trained on OPPORTUNITY and evaluated on four other datasets. Overall, CFSH-Net offers an effective and practical solution for robust sensor-based HAR under severe label scarcity and domain shift.