Yoga Pose Estimation Using Angle-Based Feature Extraction.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• The COVID-19 pandemic increased virtual yoga, highlighting challenges in maintaining correct postures.
• Accurate yoga pose feedback is crucial for effective and safe practice, especially in remote settings.

Purpose of the Study:

• To develop a computer vision and machine learning (ML) mechanism for detecting correct yoga poses.
• To provide real-time feedback for yoga practitioners using technology.

Main Methods:

• Utilized computer vision for pose estimation and angle calculation.
• Trained and tested various ML models (extremely randomized trees, logistic regression, random forest, gradient boosting, extreme gradient boosting, deep neural networks) on the Yoga-82 dataset.

Main Results:

• The extremely randomized trees model achieved the highest prediction accuracy at 91% on the test dataset.
• Other models showed varying accuracies, with deep neural networks at 85% and logistic regression performing lowest.

Conclusions:

• The extremely randomized trees model demonstrates superior predictive power for yoga pose recognition.
• The developed approach has potential for real-time feedback on low-powered smartphones, aiding home yoga practice.