COLDOG: Human Activity Recognition through Collaborative Learning for Domain Generalization
Yongtae Jeong, Seoung-Bum Kim · Journal of Korean Institute of Industrial Engineers · 2024
Human activity recognition (HAR) is a rapidly advancing field that uses wearable sensors and devices, such as smartphones, to detect and classify user activities. Despite remarkable growth, HAR models encounter challenges in generalizing to new users because of variations in user characteristics. Transfer learning and domain adaptation have been studied to address this issue, but they rely on the availability of target data. This study presents a collaborative learning for domain generalization (COLDOG) to enhance generalization performance in HAR. COLDOG improves the generalization performance of domain-specific models through collaborative learning and achieves high performance for new users through ensemble of all models. We validate the effectiveness of the proposed method through experiments on various HAR benchmark datasets.