This project is based on the pose estimation model PoseNet, ml5js and KNN classifier model
Currently this system gives around 75% accuracy which can be improve by gathering more data and retrain and fine-tune with different classifier model.The uniqueness of the project in data preparation process which is user friendly and can be inference in real-time with browser.i trained the model with 7 (seven) yoga-pose data ('Tadasana','Urdhva Hastasana','Uttanasana','Ardha Uttanasana','Chaturanga','Urdhva Mukha Svanasana','Adho Mukha Svanasana' ).If you have any query please contact me via E-mail
Pose estimation is a hot topic now-a-days. It is being used in video-surveillance system to sport analysis tasks. Some of the classical problem can be solved using pose estimation like: person count in a frame, fall detection, smart fitness tracking app etc. Basicly by using pose estimation we can observe the movement of human and take any decision. Before of deep learning arena HoG and SIFT based approach used in feature extraction. But because of CNN these feature extraction process become more accurate using lots of data.
So,Using PoseNet we get key points of human limbs. Output of keypoints is (x,y) co-ordinate value. Then using these keypoints we can determine angles of different limb of our body or can use these point in classifier model for human acitivity detection. There are some out performing model for pose estimation like: OpenPose pose estimation model which can also be inferenced in CPU.
- Python >= 3.6
- Dependencies:
pip install -r requirements.txt - Javascript
- p5.js
python manage.py runserver
- data-preparion: Data preparation files
- resize.py: code for resizing the collected images.
- video.py: code for making video from these images.
- Yoga_prediction: All the scripts for prediction with pretrain model
- index.html: Open it in broser for prediction using webcam
- Yoga_training: All the scripts for training the classifier model with pose estimation
- index.html: Open it in broser for start training using UI
- pose.js: This javascript file contain all beackend logic for pose estimation and training .
I collected yoga pose images from various sources (Flickrs,Youtube videos etc),Resize the images using resize.py . Considering these images as frame made a video using video.py by setting FPS=1.
I prepared the training procedure using UI. By clicking the button we can select current frame considering this as level. Open index.html in editor from training folder, some thing will apear like below
open index.html file from prediction folder. it will open webcam and start predicting!
