Automated Interpretable Assessment of Tai Chi Exercise Proficiency with Machine Learning, Biomechanics, and Data Visualization

Abstract

Automated systems that assess the proficiency of physical exercises and generate feedback provide significant value to individuals to improve their performance. In addition, automated systems provide practitioners and researchers a standard and rapid digital assessment tool to assist in analyzing the impact of different exercises. Tai Chi is a practice that has numerous health benefits and is of continued research interest. The “Raising the Power” Tai Chi (RTP) exercise is a classic example, with representative orchestrated movements to promote wellbeing. This thesis presents video analysis, biomechanical assessment, machine learning (ML), and data visualization methods developed and integrated to classify the RTP exercise. Random Forest and k-Nearest Neighbors classification models were created using a small (n=26) imbalanced dataset of Tai Chi videos and evaluated for 4 different scoring metrics to produce three models with greater than 70% accuracy. The decision-making criteria of these models were successfully presented in an automated data visualization that provided interpretable feedback for a prospective user. The development of this model serves as a proof of concept for the iterative development of interpretable exercise assessment models that can be started with minimal data.