Gait-Semantic Relational Knowledge Distillation for Ground Reaction Force Estimation

Area of Science:

• Biomechanics and Wearable Technology
• Human Locomotion Analysis
• Sensor Data Processing

Background:

• Wearable sensors offer portable gait analysis for healthcare and injury prevention.
• Traditional ground reaction force (GRF) measurement tools are costly and lab-bound.
• Estimating GRF from noisy wearable insole data is challenging for deep learning models.

Purpose of the Study:

• To develop a computationally efficient deep learning framework for accurate GRF estimation from wearable insole sensors.
• To leverage human-interpretable semantic information for improved gait sequence modeling.
• To address the limitations of current GRF estimation methods in real-world applications.

Main Methods:

• Proposed a Gait-Semantic Relational Knowledge Distillation (SRKD) framework.
• Incorporated gait-phase-aware semantic guidance via teacher-student knowledge distillation.
• Aligned intermediate representations with textual semantic representations for phase-specific learning.

Main Results:

• SRKD improved the GRF estimation performance of lightweight student models compared to baselines.
• Predicted GRF signals better captured waveform and phase-dependent temporal variations.
• Human-refined semantic descriptions enhanced knowledge distillation effectiveness.

Conclusions:

• The SRKD framework offers a promising approach for accurate and efficient GRF estimation using wearable insole sensors.
• Phase-aware semantic distillation effectively transfers knowledge for structured temporal dynamics in gait.
• Further validation is needed for broader applicability in diverse real-world scenarios.