Multi-User Activity Recognition in an Indoor Environment with Transformer Architectures

This paper proposes a device-free Human Activity Recognition (HAR) system, utilising Wi-Fi Channel State Information (CSI) to maintain the privacy of users in a multi-user environment. To achieve this goal, substantial annotated training data is required, which is often imbalanced with poor generalisability in complex, multi-user environments. To overcome these gaps, a hybrid deep learning approach is proposed that integrates signal pre-processing, targeted data augmentation, and a novel CNN incorporating a Transformer model. Experimental results show that the proposed model outperforms several baselines in single-user and multi-user contexts. Our findings demonstrate that combining real and augmented data significantly improves model generalisation in scenarios with limited labelled data.