COA-HAR: Exploring contrastive online test-time adaptation for wearable sensor-based human activity recognition using sensor data augmentation

초록

Human activity recognition (HAR) with wearable sensors is crucial for a wide range of applications from healthcare monitoring to sports analytics, yet challenges persist in achieving accurate and efficient predictions, particularly in real-world deployment scenarios. While deep learning techniques have improved the HAR performance, issues such as data scarcity and domain shifts from changing the user pose significant challenges. Test-time adaptation (TTA) emerges as a promising solution, offering flexibility and generality in addressing distribution shifts. However, applying TTA to sensor-based HAR has so far been limited to the adaptation of normalization layers, which leads to limited improvements. This paper introduces Contrastive Online Adaptation for HAR (COA-HAR), a novel framework leveraging self-supervised and contrastive learning to enhance model adaptability at test time. Our method is the first method for HAR to employ contrastive and self-supervised learning to directly adapt the entire model to an unlabeled setting at test time. Our contributions also include a strategy for selecting augmentations as well as an extensive study on the optimal combination of data augmentation techniques addressing the challenges of adapting these techniques to HAR. COA-HAR achieves state-of-the-art performance and we demonstrate the efficacy of leveraging pretrained models on extensive, unlabeled HAR datasets, such as the UK-Biobank dataset, to improve model initialization and adaptability. Experimental results across four datasets demonstrate the superiority of COA-HAR in HAR tasks compared to existing state-of-the-art TTA methods, showing its potential for robust and user-specific real-world applications.

키워드