Automatic feature selection for performing Unit 2 of vault in wheel gymnastics

Area of Science:

• Sports Science
• Biomechanics
• Machine Learning

Background:

• The vault event in gymnastics involves complex movements requiring precise execution.
• Execution score (E-score) deductions are critical for evaluating performance.
• Analyzing movement features objectively can improve training and judging.

Purpose of the Study:

• To develop a machine learning framework for analyzing the relationship between movement features and E-score deductions in the vaulting mounting phase.
• To identify key movement features that influence deductions during Unit 2 of the vault event.

Main Methods:

• Developed an automated system using Recurrent All-Pairs Field Transforms (RAFT) and XMem for video analysis.
• Employed OpenPose for markerless pose estimation to extract joint coordinates and angles.
• Constructed a Random Forests regression model to estimate E-score deductions based on joint angles.

Main Results:

• Achieved precise estimation of E-score deductions with a determination coefficient of 0.79.
• Identified 'time on the wheel' and 'knee angles at pike-mount' as critical features for avoiding deductions.
• Selected features aligned with judges' observations and vaulting regulations.

Conclusions:

• The proposed framework offers an objective method for analyzing gymnastics performance.
• Focusing on time on the wheel and knee angles can help gymnasts improve their vault execution.
• This approach supports skill development and adherence to competition rules.