MF-Match: A Semi-Supervised Model for Human Action Recognition.

Area of Science:

• Radar signal processing
• Machine learning for sensing
• Human-computer interaction

Background:

• Human action recognition (HAR) using radar offers privacy-preserving, noncontact sensing independent of lighting.
• Deep learning models for HAR require large labeled datasets, which are scarce for radar data.
• This data scarcity hinders the advancement of radar-based HAR technology.

Purpose of the Study:

• To address the challenge of limited labeled data in radar-based HAR.
• To propose a semi-supervised learning algorithm that leverages unlabeled radar data.
• To enhance the accuracy and robustness of human action recognition models.

Main Methods:

• Developed MF-Match, a semi-supervised algorithm to generate pseudo-labels for unsupervised radar data.
• Integrated contrastive learning to refine pseudo-label quality and minimize mislabeling impact.
• Applied the algorithm to extract embedded human behavioral information from radar signals.

Main Results:

• Achieved high action recognition accuracies of 86.69% and 91.48% on two radar datasets.
• Demonstrated effectiveness using only 10% of labeled data, significantly reducing data requirements.
• Validated the algorithm's capability to improve HAR performance despite limited labeled data.

Conclusions:

• MF-Match effectively overcomes the data scarcity problem in radar-based HAR.
• The proposed method enhances HAR accuracy by utilizing unlabeled data and contrastive learning.
• This approach significantly advances the feasibility of large-scale deep model-based HAR systems.