Related Experiment Video
Updated: Jul 17, 2026
06:37
Artificial Intelligence-Based System for Detecting Attention Levels in Students
Published on: December 15, 2023
Few-Shot Learning for Cross-Domain Human Activity Recognition Using Wearable Sensors
IEEE Journal of Biomedical and Health Informatics
|July 15, 2026
Summary
A new network, CFSH-Net, enables accurate human activity recognition (HAR) with limited data by adapting to new sensors and users. This lightweight model overcomes challenges in health monitoring and smart homes.
Area of Science:
- Pervasive computing and artificial intelligence.
- Biomedical engineering and health informatics.
Background:
- Sensor-based human activity recognition (HAR) is vital for health monitoring and smart environments.
- Deep learning for HAR faces limitations due to scarce labeled data and domain shifts (user, sensor, scenario variability).
Purpose of the Study:
- To propose a novel lightweight cross-domain few-shot (FS) sensor-based HAR network (CFSH-Net).
- To enable robust HAR with limited labeled samples and adapt to unseen activities and domain shifts without fine-tuning.
Main Methods:
- Utilized wavelet-driven depth-wise bilateral decomposition (WDBD) as a feature extractor.
- Developed a network architecture for cross-domain few-shot learning in sensor-based HAR.
- Evaluated on six public benchmarks: OPPORTUNITY, PAMAP2, UCI-HAR, WISDM, WISDM2019, and USC-HAD.
Main Results:
- Achieved 85.22% accuracy for unseen activities on PAMAP2 and 88.66% on OPPORTUNITY (5-way 5-shot).
- Reached 79.94% average accuracy in cross-position recognition.
- Demonstrated strong cross-user generalization and stable cross-dataset transfer.
Conclusions:
- CFSH-Net provides an effective and practical solution for robust sensor-based HAR.
- The proposed method addresses severe label scarcity and domain shift challenges in real-world applications.
