Selective Correlation Based Knowledge Distillation for Ground Reaction Force Estimation

Area of Science:

• Biomechanics
• Sensor Technology
• Machine Learning

Background:

• Wearable sensors offer portable human gait analysis, crucial for healthcare and sports.
• Traditional ground reaction force (GRF) measurement is limited by expensive lab equipment.
• Insole sensors for GRF estimation suffer from noise and accuracy issues, while deep learning demands high computational resources.

Purpose of the Study:

• To develop a resource-efficient deep learning method for accurate GRF estimation from wearable insole sensor data.
• To address the limitations of noise, interference, and computational demands in portable gait analysis.
• To enhance the interpretability and applicability of deep learning models for real-time GRF analysis.

Main Methods:

• Proposed Selective Correlation Based Knowledge Distillation (SCKD) framework for GRF estimation.
• Utilized selected features considering temporal characteristics for correlation map extraction and knowledge transfer.
• Examined various teacher-student architectures and training approaches with data from different walking speeds and window sizes.

Main Results:

• SCKD demonstrated superior performance in estimating GRF compared to existing methods.
• The proposed distillation framework generated compact, accurate deep learning models.
• Experimental results confirmed the effectiveness of SCKD across different gait conditions.

Conclusions:

• SCKD provides a reliable and resource-efficient solution for human gait analysis using wearable insole sensors.
• The method enhances accuracy and interpretability in GRF estimation.
• Enables real-time gait analysis on portable devices, advancing healthcare and sports applications.