Estimating Ground Reaction Forces From Inertial Sensors.

Area of Science:

• Biomechanics
• Sports Science
• Machine Learning

Background:

• Ground reaction forces (GRFs) characterize mechanical loading during running, crucial for identifying injury risks.
• Current state-of-the-art methods for GRF estimation using LSTMs are computationally intensive and lack transparency.
• Inertial measurement units (IMUs) offer a portable solution for collecting running data.

Purpose of the Study:

• To evaluate lightweight machine learning approaches for estimating GRFs and biomechanical variables from IMU data.
• To compare the accuracy and efficiency of novel lightweight methods against traditional deep learning models.
• To assess the impact of personalized data on estimation accuracy.

Main Methods:

• Proposed SVD Embedding Regression (SER), a novel lightweight method.
• Compared SER and k-Nearest-Neighbors (KNN) regression against LSTMs.
• Utilized IMU data (acceleration, angular velocity) from sacrum and shanks in various experimental scenarios.

Main Results:

• Lightweight methods (SER, KNN) demonstrated comparable or superior accuracy to LSTMs for GRF and biomechanical variable estimation.
• Personalized training data significantly reduced estimation errors, especially for biomechanical variables with lightweight methods.
• Explored sensor locations and data combinations for optimal performance.

Conclusions:

• Lightweight machine learning models are effective for estimating running biomechanics from IMU data.
• SER and KNN provide efficient and accurate alternatives to complex deep learning models.
• Personalized data enhances the performance of lightweight models for practical applications in sports science and injury prevention.