FSID: a novel approach to human activity recognition using few-shot weight imprinting

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Human Activity Recognition (HAR) is crucial for healthcare but challenged by limited labeled data and poor generalization in conventional deep learning.
• Existing models struggle with scarce, imbalanced, or novel activities, hindering personalized and dynamic healthcare applications.

Purpose of the Study:

• To propose Few-Shot Imprinted DINO (FSID), a novel framework for HAR in low-data regimes.
• To address the limitations of conventional deep learning models in real-world scenarios with insufficient or imbalanced datasets.

Main Methods:

• Converted time-series sensor data (EMG, IMU) into spectrogram images via Short-Time Fourier Transform.
• Utilized a pre-trained DINO (Distillation with No Labels) model as a feature extractor for transferable representations.
• Employed Few-Shot learning with weight imprinting to create classification prototypes without iterative fine-tuning.

Main Results:

• FSID achieved up to 55.47% accuracy on the HuGaDB dataset and 35.81% on the LARa dataset with only 20 novel samples.
• Demonstrated superior performance compared to baseline models across various configurations in low-data settings.
• Spectrograms proved effective and computationally efficient for time-frequency representation.

Conclusions:

• FSID offers an effective solution for HAR in low-data regimes, enhancing applicability in personalized healthcare.
• The framework successfully leverages self-supervised learning and weight imprinting for robust activity recognition with minimal data.
• The findings highlight the potential of FSID for real-world applications requiring efficient and accurate HAR from gait sensory data.