Efficient Human Activity Recognition on Smartwatch Sensors Using Knowledge Distillation
1
Issued Date
2026-01-01
Resource Type
ISSN
23673370
eISSN
23673389
Scopus ID
2-s2.0-105021817192
Journal Title
Lecture Notes in Networks and Systems
Volume
1675 LNNS
Start Page
681
End Page
697
Rights Holder(s)
SCOPUS
Bibliographic Citation
Lecture Notes in Networks and Systems Vol.1675 LNNS (2026) , 681-697
Suggested Citation
Hnoohom N., Mekruksavanich S., Jitpattanakul A. Efficient Human Activity Recognition on Smartwatch Sensors Using Knowledge Distillation. Lecture Notes in Networks and Systems Vol.1675 LNNS (2026) , 681-697. 697. doi:10.1007/978-3-032-07986-2_42 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/114654
Title
Efficient Human Activity Recognition on Smartwatch Sensors Using Knowledge Distillation
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
This study proposes a computationally efficient human activity recognition (HAR) framework on smartwatches by leveraging knowledge distillation techniques. It addresses the challenge of deploying deep learning models on resource-constrained wearable devices through the use of the WISDM-HARB dataset, which contains accelerometer and gyroscope data collected from both smartphones and smartwatches across 18 human activities. This approach transfers knowledge from a sophisticated teacher network comprising 12,648,664 parameters to a lightweight student model containing only 66,002 parameters. Despite this significant reduction in model size and complexity, the student network achieves 94.98% classification accuracy—closely matching the teacher model’s 95.14%—while reducing computational overhead by 98.48%, from 411.8 million to 6.2 million FLOPs, and model size by 99%. Extensive experiments across various hyperparameters show that lower values of the distillation coefficient (α = 0.1–0.2) and moderate temperature settings (T = 2–5) yield the best performance. These findings demonstrate that knowledge distillation can effectively compress deep HAR models without significant loss in accuracy, offering a practical solution for real-time activity recognition on low-power smartwatch devices.