A meta learning framework for few shot personalized gait cycle generation and reconstruction

Area of Science:

• Biometrics
• Machine Learning
• Robotics

Background:

• Human gait is a complex biometric with high variability, posing challenges for deep learning personalization.
• Current deep learning models often require extensive subject-specific data for accurate gait generation and reconstruction.

Purpose of the Study:

• To introduce MetaGait, a novel framework utilizing meta-learning for efficient personalization of gait models.
• To enable accurate gait analysis and reconstruction from limited data using few-shot learning.

Main Methods:

• MetaGait employs a Model Agnostic Meta Learning (MAML) strategy with a Temporal Convolutional Network (TCN).
• A base model is trained on diverse gait tasks from the Human Gait Database (HuGaDB), optimizing for rapid adaptation.
• Few-shot learning (1-shot and 5-shot) is used to adapt the model to specific walking conditions with minimal gait cycles.

Main Results:

• MetaGait significantly outperforms conventional models in few-shot gait cycle generation and reconstruction, as measured by MSE and DTW.
• Qualitative evaluations demonstrate MetaGait's ability to produce natural, subject-specific gait patterns.
• The framework achieves accurate reconstructions even with sparse input data.

Conclusions:

• MetaGait offers a practical solution for gait personalization by drastically reducing data requirements.
• The meta-learning approach enhances the adaptability of gait models for real-world applications in robotics and clinical settings.