Predicting Athlete Ground Reaction Forces and Moments From Spatio-Temporal Driven CNN Models

Area of Science:

• Biomechanics
• Machine Learning
• Sports Analytics

Background:

• Accurate 3-D ground reaction forces and moments (GRF/Ms) prediction is crucial for on-field biomechanical analysis.
• Current methods rely on ground-embedded force plates, limiting real-world applications.
• Previous approaches, like partial least squares (PLS), have limitations in accuracy and scope.

Purpose of the Study:

• To develop a deep learning model for predicting 3-D GRF/Ms from marker-based motion capture data.
• To improve upon existing partial least squares (PLS) methods for GRF/Ms prediction.
• To enable biomechanical analysis outside of laboratory settings.

Main Methods:

• Utilized five Convolutional Neural Network (CNN) models for multivariate regression of 3-D GRF/Ms.
• Employed tactical feature engineering to compress space-time data.
• Fine-tuned three pretrained CNNs, including a CaffeNet model derived from ImageNet.

Main Results:

• The CaffeNet model demonstrated the highest correlation with ground truth GRF/Ms (R² = 0.9881, R² = 0.9715).
• Achieved low prediction errors ([Formula: see text] = 4.31% and [Formula: see text] = 7.04%).
• Validated the effectiveness of CNNs for real-world multivariate regression in biomechanics.

Conclusions:

• Deep learning, particularly CNNs, offers a powerful tool for predicting 3-D GRF/Ms from motion capture.
• This approach significantly reduces reliance on laboratory force plates for biomechanical analysis.
• The findings have practical implications for spatio-temporal sports analytics and performance optimization.