Yoga Pose Estimation and Tracking

Human pose keypoint detection and estimation for yoga

Overview

This project aimed to develop a real-time yoga pose detection and correction system using computer vision (CV) techniques and statistical modeling. The system leveraged MediaPipe and OpenCV to capture and process image data for accurate pose estimation, joint angle computation, and feedback generation. Worked with Prof. Kunal Korgaonkar Assistant Professor at the CSIS Dept. of BITS Pilani, K K Birla Goa Campus in the CSIS Robotics Lab. The approach involved computing joint angles from processed images, comparing them with predefined statistical ranges using Z-score analysis, and suggesting corrections to improve pose accuracy.

Project Contributions:

  • Developed a real-time yoga pose tracking system integrating depth sensors and computer vision techniques.
  • Implemented MediaPipe for rapid pose detection and tracking using lightweight perception pipelines.
  • Extracted joint angles from processed images and structured datasets for analysis.
  • Curated a statistical dataset with Z-distribution-based angle ranges for each yoga pose.
  • Computed Z-scores to compare real-time poses with mean values and assess deviation.
  • Designed feedback algorithms to suggest pose corrections dynamically.
  • Confidence in pose measurements fall within natural biomechanical limits across individuals.
  • Implemented confidence thresholds:
    • Minimum Detection Confidence: Ensuring reliable pose identification.
    • Maximum Tracking Confidence: Maintaining robust real-time tracking.

Technical Highlights:

  • MediaPipe Framework: Enabled rapid prototyping and deployment of computer vision pipelines.
  • OpenCV for Image Processing: Extracted skeletal features and computed pose angles.
  • Statistical Z-Score Analysis: Compared joint angles to pre-defined pose accuracy thresholds.
  • Real-Time Pose Extraction & Classification: Used 3D angle calculations to label poses dynamically.
  • Confidence Interval Computation: Improved pose consistency across different individuals.
  • Error Handling & Adaptive Tuning: Made the model resilient to minor pose variations.
Detected Landmarks

Impact and Future Scope:

This system enhances yoga training, fitness applications, and physiotherapy by providing automated pose correction and feedback. The work can be extended to:

  • Personalized AI-driven yoga assistants with real-time coaching.
  • Rehabilitation & physiotherapy monitoring for patient recovery.
  • Integration with Augmented Reality (AR) for interactive fitness solutions.

GitHub Repository and Project Report: Yoga Pose Detection System