UniHAR: A unified hierarchical transformer with multi-scale temporal modeling for wearable human activity recognition

UniHAR is a lightweight hierarchical multi-scale transformer based on wearable sensor data that has to be able to model time well; be able to recognize human activity across different datasets; and do so in real-time. There are several problems with what currently exists. UniHAR uses four major components: universal channel harmonization; adaptive multi-scale patch embedding; scale-adaptive Fourier cross-attention; factorized temporal-channel attention. The components allow for the ability to use various sensors and enables the low-complexity multi-scale fusion of features. UniHAR achieved an accuracy of 96.35%; 94.45%; 96.44%; and 99.23% on UCI-HAR, Opportunity, Daphnet, and PAMAP2 datasets, while being very lightweight and having high-speed inference capabilities; therefore, UniHAR is suitable for real-time healthcare applications.