Geometric Feature Relationship-Based Knowledge Distillation for Ground Reaction Force Estimation

Area of Science:

• Biomechanics
• Sensor Technology
• Machine Learning

Background:

• Wearable sensors offer portable gait analysis for healthcare and injury prevention.
• Traditional ground reaction force (GRF) measurement is lab-bound and costly.
• Existing deep learning models for GRF estimation require significant computational resources, limiting real-time use.

Purpose of the Study:

• To develop a knowledge distillation method for creating compact, efficient deep learning models for GRF estimation from wearable insole sensors.
• To address limitations in current distillation techniques that use fixed similarity operators and fail to capture shared nonlinear feature geometry.

Main Methods:

• Propose Geometric Feature Relationship-Based Knowledge Distillation (GFKD) for GRF estimation.
• Introduce a teacher-student shared assistant module (TSM) for intermediate representation alignment.
• TSM computes and enforces nonlinear geometric relations between teacher and student models.

Main Results:

• GFKD-distilled student models outperform existing baselines in GRF estimation accuracy.
• The proposed method demonstrates robustness in both accuracy and resource-efficiency.
• Achieved superior performance compared to recent knowledge distillation techniques.

Conclusions:

• GFKD provides a geometry-aware supervision method, improving knowledge transfer.
• The approach enables accurate and resource-efficient GRF estimation using wearable sensors.
• GFKD is a promising solution for real-time, on-device gait analysis applications.