Towards Customizable Foundation Models for Human Activity Recognition with Wearable Devices

Foundation models have achieved remarkable success across various domains by learning general representations from raw data, offering a promising paradigm for diverse applications. This concept holds great potential for advancing human activity recognition (HAR), particularly in overcoming challenges associated with collecting large-scale labeled datasets. However, the dynamic nature of HAR tasks, characterized by diverse sensing devices and activity types, results in fragmented datasets that question the feasibility of applying foundation model to this domain. In this work, we propose a novel foundation model training framework that effectively leverages heterogeneous datasets through a two-stage training strategy: (1) self-supervised learning to extract cross-domain sensor patterns, followed by (2) multi-task learning to align representations with semantic contexts. The effectiveness of the trained foundation model is demonstrated through extensive downstream experiments, with the superior fine-tuning performance across various modalities and input configurations---achieving the highest performance metric in 10 out of 12 settings---further validating the robustness and adaptability. While our model shows a performance gap compared to foundation models pre-trained on large-scale or high-quality data in zero- and few-shot scenarios, its competitive results with a more flexible architecture demonstrate the efficiency and potential of our training strategy for HAR foundation models.