Exercise quantification from single camera view markerless 3D pose estimation.

Area of Science:

• Biomechanics
• Sports Science
• Human Movement Analysis

Background:

• Visual assessment of athlete movement quality is often inaccurate.
• Existing motion capture technology is expensive and lacks widespread adoption.
• There is a need for accessible tools to measure joint angles and movement metrics.

Purpose of the Study:

• To define pose landmarks and metrics for assessing common strength-building exercises.
• To develop rules for computing these metrics from video-based key-point detection.
• To evaluate the accuracy of these metrics using a 3D human pose model.

Main Methods:

• Defined a set of pose landmarks for calculating joint angles.
• Developed algorithms to compute metrics for exercises like squats, deadlifts, and jumps.
• Validated the accuracy against VICON motion capture data using a Strided Transformer 3D pose model.

Main Results:

• Mathematically defined metrics were sufficient for exercise assessment.
• Root mean square angle errors were within 10° for key metrics like shin, knee, and hip angles.
• Errors within 15° were observed for shoulder flexion and ASIS asymmetry in specific exercises.

Conclusions:

• A unique set of key-points and metrics for 3D human pose-based exercise assessment has been established.
• The Strided Transformer 3D pose model demonstrated high accuracy in predicting these metrics from single-camera video.
• Further research is needed to assess the generalization capabilities of the model.