FSID: A Novel Approach to Human Activity Recognition Using Few-Shot Weight Imprinting

Accurate recognition of activities from gait sensory data is essential for healthcare and wellness monitoring applications. This paper proposes Few-Shot Imprinted DINO (FSID), a new approach for Human Activity Recognition (HAR) in scarce-data settings, combining Few-Shot learning with weight imprinting. The methodology transforms gait sensory signals into spectrograms using the Fourier Transform, enabling the use of deep learning techniques like transfer learning to refine activity classification. FSID extracts meaningful features from generated spectrograms using the DINO model, integrating transfer learning with Few-Shot learning and weight imprinting to effectively classify uncommon or novel activities. Extensive experimentation with diverse datasets, such as HuGaDB and LARa, demonstrates that FSID outperforms current methods, achieving high accuracy and robustness even with limited labeled data. These results confirm that spectrogram representations, together with effective Few-Shot learning integration, enhance the model's scalability and adaptability, making FSID suitable for healthcare applications where data collection may be challenging.