/yoga-pose-estimation

Primary LanguageJupyter NotebookApache License 2.0Apache-2.0

yoga-pose-estimation

This project is in reference to the papers:

  • "Real-time Yoga recognition using deep learning" by Santosh Kumar Yadav, Amitojdeep Singh, Abhishek Gupta & Jagdish Lal Raheja
  • "Yoga-82: A New Dataset for Fine-grained Classification of Human Poses" by Manisha Verma, Sudhakar Kumawat, Yuta Nakashima, Shanmuganathan Raman

Requirements:

  1. Videos used can be found here: https://www.researchgate.net/deref/https%3A%2F%2Farchive.org%2Fdetails%2FYogaVidCollected
  2. Video frames used can be found here: https://drive.google.com/drive/folders/1S6r6YH8upulktnfDbEn5DRxwQgqeXZN3?usp=sharing
  3. Steps to install OpenPose on the system can be found here: https://github.com/CMU-Perceptual-Computing-Lab/openpose
    • to get 25 joint data, model used should be: BODY_25
    • to get 18 joint data, model used should be: COCO

Few details of Video dataset:

  1. 6 classes, 15 participants
  2. Duration of each video: 45+ seconds with 30 frames per second
  3. Surroundings: Indoor environment
  4. Total videos: 88
  5. Combined time: 1 hour, 6 minutes, and 5 seconds at 30 fps
  6. Data split ratio: 60:20:20 (approx.)

Results:

Results of the CNN-LSTM model on the 18 joint data of OpenPose are given below:

  • Train Accuracy: 99.95%
  • Validation Accuracy: 99.60%
  • Test Accuracy: 99.87%