MetaRadarHAR: A Radar-Based Human Activity Recognition Methodology Using Metric-Based Meta-Learning

Radar-based human activity recognition (HAR) methods are often constrained by the limited availability of open-source datasets and the challenges of collecting large-scale radar data. To address this issue, we propose a novel methodology that leverages metric-based meta-learning to alleviate the need for large-scale datasets, and utilizes a two-stage training strategy to enhance model accuracy and generalization. Specifically, after data preprocessing and getting radar time-range (TR) feature maps, a feature extractor is used to embed the features, and activity classification is performed based on the cosine similarity between the embeddings and the class prototypes. To improve efficiency, we introduce a lightweight network based on structural reparameterization as the feature extractor, which uses only 1/25 of the computational resources and 60% of the parameters compared to the commonly used ResNet-12. Cross-validation experiments on the public dataset IURHA2023-TR1 and the self-collected dataset demonstrate the effectiveness of our methodology, achieving an average recognition accuracy of 90.57% for 20 activities using only 30 training samples.