Location independent human activity recognition using self-training CSI-based techniques for wireless sensor networks
Human activity recognition (HAR) using WiFi is applied across various domains ranging from smart environments, the Internet of Things (IoT) and immersive virtual gaming. The environmental effects of WiFi sensing lie in its susceptibility to variations in physical surroundings, which influence signal...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | en |
| Published: |
Institute of Electrical and Electronics Engineers Inc.
2025
|
| Online Access: | http://eprints.utem.edu.my/id/eprint/29306/2/01298100720251458.pdf http://eprints.utem.edu.my/id/eprint/29306/ https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10988621&tag=1 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Human activity recognition (HAR) using WiFi is applied across various domains ranging from smart environments, the Internet of Things (IoT) and immersive virtual gaming. The environmental effects of WiFi sensing lie in its susceptibility to variations in physical surroundings, which influence signal strength and accuracy in detecting human activity. Innovative solutions are needed to meet these demands, such as activity-adapted learning for seamless feature transfer
and recognition across various locations, reducing the reliance on extensive training datasets. This work proposes a framework incorporating a confidence threshold to filter unreliable samples, a progressive self-training strategy to integrate unlabeled data, and a weighted self-training approach to counter class imbalance. The proposed model explores HAR and its improved performance by integrating self-training techniques. This work enhances HAR by reconciling self-training’s potential with challenges and
offering practical insights for reliable activity recognition within wireless sensor networks. The results of experiments show that the self-training method, which uses channel state information based features to train the model with unlabeled data, is up to 97.5% accurate. Additionally, experiments using HAR datasets validate the proposed method and displays performance improvements over baselines. |
|---|