TinyHAR-Net: Design and Implementation of a TinyML-Based Human Activity Recognition Framework on STM32
IEEE Embedded Systems Letters · 2026-08 · 18 (4) · 308-311
DOI: 10.1109/les.2025.3639944 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Embedded Systems Letters, 2026-08. Verfasst von Imran Hossan und 2 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref) und nennt 13 Literaturangaben. Gefördert von: University Grants Commission (UGC), Bangladesh, Rajshahi University of Engineering and Technology.
Abstract (Originalsprache)
Deploying high-performing and resource-efficient human activity recognition (HAR) models on microcontrollers (MCUs) remains a significant challenge due to severe limitations in memory and computational capacity. While methods, such as quantization and pruning, can reduce model size, they mostly degrade the model’s performance. In this letter, we introduce TinyHAR-Net, a novel lightweight hybrid model designed specifically for efficient, real-time HAR on embedded devices, without needing those extra optimization steps. The model combines depthwise separable convolutions with a single-head Linformer self-attention mechanism to effectively capture both local and global features in sensor data. We evaluated the model on three public datasets, including WISDM, MotionSense, and MHEALTH, and deployed it on an STM32F446RE MCU. TinyHAR-Net achieved ${F}1$ scores of 96.75% on WISDM, 97.73% on MotionSense, and 99.76% on MHEALTH, with corresponding flash memory usage of 43.78, 47.16, and 51.30 kB and inference latency of 41.31, 46.37, and 51.57 ms, respectively. RAM consumption was consistently 30.12 kB across all datasets. Without relying on compression, TinyHAR-Net provides a reliable and efficient solution for real-time HAR on devices with limited resources.
Förderung
- University Grants Commission (UGC), Bangladesh
- Rajshahi University of Engineering and Technology
Zitieren
APA
Hossan, I., Mary, M. N. J., & Motin, M. A. (2026). TinyHAR-Net: Design and Implementation of a TinyML-Based Human Activity Recognition Framework on STM32. IEEE Embedded Systems Letters, 18(4), 308-311. https://doi.org/10.1109/les.2025.3639944
Vancouver
Hossan I, Mary MNJ, Motin MA. TinyHAR-Net: Design and Implementation of a TinyML-Based Human Activity Recognition Framework on STM32. IEEE Embedded Systems Letters. 2026;18(4):308-311. doi:10.1109/les.2025.3639944
BibTeX
@article{Hossan2026,
author = {Hossan, Imran and Mary, Mst. Nusratul Jannat and Motin, Mohammod Abdul},
title = {TinyHAR-Net: Design and Implementation of a TinyML-Based Human Activity Recognition Framework on STM32},
journal = {IEEE Embedded Systems Letters},
year = {2026},
volume = {18},
number = {4},
pages = {308--311},
doi = {10.1109/les.2025.3639944}
}
EUVIMED