DefinePK

Abstract

Human Activity Recognition has become inevitable in wearable technology in healthcare and fitness tracking applications. Precise recognition of daily routine human activities using wearable sensors remains a challenging task due to sensor modality, placement, and environmental factors. Indeed, Machine Learning and Deep Learning models have become robust in human activity recognition, yet these face numerous challenges in precisely recognizing the daily human activities. One major challenge is the variability in sensors readings where the same activity may lead to different sensor readings, intra-class variability, as different people perform the same activity differently or use different devices. To address this, we propose a multi-head transformer model with a multi-head attention mechanism that explicitly handles the intra-class variability. We demonstrate that the multi-head Transformer model exhibits enhanced robustness and performs better even in the face of such variability, although variability is a major problem in the popular PAMAP2 dataset, since it directly impacts the performance of deep learning models.