hungmeomeo / posture-detection Goto Github PK

A lightweight posture classifier using frontal camera images. Provide API for frontend using client-server method with framework Flask of Python.

• OpenCV-Python
• Mediapipe
• scikit-learn
• win10toast

Run 'api.py', after the Flask finished preparation, frontend framework can call api through socket with port 5000. The API receives a list of detected keypoints and returns the prediction according to them.

Run datacollect.py. Then switch to an 'okay' posture while pressing the key B. Then switch to an 'bad' posture while pressing the key C. Then switch to an 'terrible' posture while pressing the key D. Data instance count will be shown on the terminal. 400 instances for each category works well for me. After collecting, press Q to exit the program and a posture_data.npy file will be saved.

Run trainclassifer.py to train the model, a simple scikit-learn LogisticRegression classifier. The model will be saved as model.sav

In addition, train the model through this Colab Notebook for better accuracy: ML Pipeline

To see how well your model is, run predict_demo.py. The classified posture will be shown as text on screen.

Run predict_nofity.py to use the actual posture remainder program. The classifier will run in the background and pop up a Windows notification in the lower right corner whenever you are in a bad sitting posture.