Abstract
Accurate recognition of human activities from gait sensory data plays a vital role in healthcare and wellness monitoring. However, conventional deep learning models for Human Activity Recognition (HAR) often require large labeled datasets and extensive training, which limits their effectiveness in real-world scenarios with scarce or imbalanced data. These models also struggle to generalize to rare or unseen activities, making them less suitable for dynamic and personalized healthcare settings. This paper proposes Few-Shot Imprinted DINO (FSID), a novel framework for HAR in low-data regimes, combining Few-Shot learning with weight imprinting and a self-supervised vision transformer, DINO (Distillation with No Labels). The FSID pipeline begins by converting raw time-series sensor data (e.g., EMG and IMU signals) into spectrogram images using the Short-Time Fourier Transform. Spectrograms were selected due to their superior performance and computational efficiency compared to alternative time-frequency representations like scalograms. These spectrograms are then passed to a pre-trained DINO model, which is used as a feature extractor due to its ability to learn global, transferable representations without requiring manual labels. Within the Few-Shot learning framework, we compute mean feature embeddings from limited support samples, and use weight imprinting to directly form classification prototypes without iterative fine-tuning. Extensive experiments on HuGaDB and LARa publicly available datasets demonstrate the effectiveness of FSID. The approach achieves up to 55.47% accuracy on HuGaDB and 35.81% on LARa with only 20 novel samples, outperforming baseline models across all configurations.