PACL+: Online continual learning using proxy-anchor and contrastive loss with Gaussian replay for sensor-based human activity recognition

Abstract

The increasing prevalence of wearable sensors and devices necessitates the development of Human Activity Recognition (HAR) systems that maintain accurate performance over time. Traditional HAR models, which rely on offline supervised training, struggle to adapt to the dynamic nature of real-world environments, leading to performance degradation due to catastrophic forgetting when new activities or users are introduced. In this paper, we propose a novel continual learning method (PACL+) that integrates Proxy Anchor loss, contrastive learning, and Gaussian replay to mitigate catastrophic forgetting in HAR systems and improve HAR performance. Unlike previous approaches, PACL+ effectively handles the introduction of both new activities and new users in incremental learning steps, addressing real-world challenges such as severe subject-wise class imbalance and user-dependent learning. To improve efficiency, we introduce Gaussian replay, a memory-efficient strategy that selects representative examples for rehearsal, further stabilizing the learning process. We evaluate PACL+ on three benchmark HAR datasets under realistic continual learning scenarios with varying sampling rates and diverse class distributions. Experimental results demonstrate that PACL+ significantly outperforms existing state-of-the-art methods, achieving higher accuracy and F1 scores while preserving performance on previously learned activities.