Machine Learning and Deep Neural Network Architectures for 3D Motion Capture Datasets.

Area of Science:

• Biomechanics
• Machine Learning
• Data Science

Background:

• Biomechanical movement data are complex, multivariate time-series.
• Applying deep learning, particularly convolutional neural networks, to 3D movement data is challenging due to nonlinear correlations.
• Deep neural networks can potentially reduce manual data processing and feature engineering in biomechanics.

Purpose of the Study:

• To determine the most appropriate classification techniques for biomechanical movement data.
• To establish baseline performance metrics for 3D joint center trajectory classification using traditional machine learning.
• To facilitate robust comparisons between various classifier architectures.

Main Methods:

• A framework and dataset were developed for comparing classification techniques.
• The study involved 416 athletes across various levels and sports.
• Thirteen non-sport-specific movements were analyzed, focusing on 3D joint center trajectories.

Main Results:

• Baseline performance of traditional machine learning techniques for 3D joint center trajectory classification is presented.
• A comprehensive dataset supporting classifier architecture comparison has been established.
• Deep neural networks tailored for time-series data are under evaluation.

Conclusions:

• Traditional machine learning provides a performance baseline for biomechanical movement classification.
• Deep neural networks are being evaluated for their potential to outperform traditional methods on complex biomechanical data.
• The developed framework and dataset enable rigorous evaluation of advanced classification techniques in sports biomechanics.