Energy-efficient activity and sleep recognition using edge computing and wearable devices

Energy-efficient activity and sleep recognition using edge computing and wearable devices

Paper in proceedings outside WoS and Scopus

Recent advancements in human activity recognition (HAR) have facilitated its integration into diverse applications, including the monitoring of vulnerable individuals. This work presents an edge computing-based framework for activity classification and sleep detection, emphasizing safety, energy efficiency, and minimized data transmission. Subsequently, dedicated algorithms for HAR and sleep detection were designed and implemented. Five HAR models and four sleep detection models were trained using selected datasets. The proposed system performs feature extraction directly on the edge device, significantly reducing communication overhead. The feature extraction methods are designed in a way that also reduces energy consumption. Results demonstrate that statistical feature computation on the edge device reduces energy consumption by around 5× compared to the on-device classification method, and data transmission volume by approximately 35× compared to a system transmitting raw sensor data. The random forest classifier achieved the highest HAR accuracy of 96.9%, while for sleep detection, the random forest reached an accuracy of 91.5%. The proposed architecture offers an effective trade-off between classification accuracy, energy usage, and communication cost, making it suitable for deployment in constrained edge environments.

Recent advancements in human activity recognition (HAR) have facilitated its integration into diverse applications, including the monitoring of vulnerable individuals. This work presents an edge computing-based framework for activity classification and sleep detection, emphasizing safety, energy efficiency, and minimized data transmission. Subsequently, dedicated algorithms for HAR and sleep detection were designed and implemented. Five HAR models and four sleep detection models were trained using selected datasets. The proposed system performs feature extraction directly on the edge device, significantly reducing communication overhead. The feature extraction methods are designed in a way that also reduces energy consumption. Results demonstrate that statistical feature computation on the edge device reduces energy consumption by around 5× compared to the on-device classification method, and data transmission volume by approximately 35× compared to a system transmitting raw sensor data. The random forest classifier achieved the highest HAR accuracy of 96.9%, while for sleep detection, the random forest reached an accuracy of 91.5%. The proposed architecture offers an effective trade-off between classification accuracy, energy usage, and communication cost, making it suitable for deployment in constrained edge environments.

human activity recognition, sleep detection, edge computing, wearable devices, person monitoring

human activity recognition, sleep detection, edge computing, wearable devices, person monitoring

ROSA, M.; BURGET, R.; HUBÁLEK, J.

2026

03.11.2025

IEEE

2025 17th International Congress on Ultra Modern Telecommunications and Control Systems and Workshops (ICUMT)

211

217

7

https://ieeexplore.ieee.org/document/11268695