hunchback-detection 驼背检测

A lightweight sitting posture classifier using frontal camera images. You can optionally send a Windows notification to remind you. This is what I used personally to help correct my sitting posture. The training data describes my own behavior, you may need to collect your own data and train your own model (takes about 2 minutes).

Dependencies

opencv-python
mediapipe
scikit-learn
win10toast (for poping Windows notification)

Data Collection

Run datacollect.py, after seeing your images, maintaining an 'excellent' posture while pressing the key A. 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 ESC to exit the program and a posture_data.npy file will be saved.

Model Training

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

Demo

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

Actual Usage

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.