PhD Forum: Domain Adaptation Approach for WiFi-Based Human Activity Recognition

In recent years, WiFi-based Human Activity Recognition (HAR) has gained attention for its usefulness across various applications. However, models trained in one environment often suffer from reduced accuracy in new settings due to domain shift. Additionally, many existing HAR models are computationally heavy, making them unsuitable for resource-constrained devices. To address these challenges, we propose a lightweight HAR method and a feature augmentation-based domain adaptation approach (FA-DA), using an efficient time-series feature extractor called MiniROCKET. We first trained RidgeCV and CNN classifiers on MiniROCKET features, achieving average F1-scores of 84.85% and 80.45%, respectively, in a single environment. However, performance dropped to less than 20% when tested in a new room. Our FA-DA method significantly improved crossenvironment performance, raising F1-scores to 80.35% for the RidgeCV classifier and 76.90% for the CNN classifier without retraining. This demonstrates the effectiveness of our approach for real-world HAR scenarios where environments vary.