Ground Reaction Force Estimation via Time-aware Knowledge Distillation.

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Human gait analysis is crucial for healthcare, rehabilitation, and diagnostics.
• Instrumented treadmills are accurate but costly and non-portable for measuring ground reaction forces (GRF).
• Wearable insole sensors offer portability but face accuracy and noise challenges, while deep learning solutions are resource-intensive.

Purpose of the Study:

• To propose an efficient and accurate method for estimating GRF from wearable insole sensor data.
• To address the limitations of current GRF measurement techniques for portable gait analysis.
• To develop a computationally inexpensive deep learning framework suitable for real-time applications.

Main Methods:

• Developed a Time-aware Knowledge Distillation framework to leverage temporal and feature similarity.
• Utilized lightweight 'student' models trained via distillation from more complex 'teacher' models.
• Evaluated performance against instrumented treadmill GRF measurements across various walking speeds and model architectures.

Main Results:

• The proposed Time-aware Knowledge Distillation framework significantly outperformed existing baselines in GRF estimation accuracy.
• Distilled lightweight models achieved excellent accuracy and reliability comparable to more complex methods.
• The framework demonstrated a substantial reduction in required training parameters, enhancing efficiency.

Conclusions:

• Time-aware Knowledge Distillation provides a data- and resource-efficient solution for accurate GRF estimation from wearable sensors.
• This approach enables practical, real-time human gait analysis in portable and diverse settings.
• The framework enhances the feasibility of wearable sensor technology for widespread healthcare and rehabilitation applications.