A lightweight posture tracking app built during the AI Research Hackathon at UNC Charlotte. Uses MediaPipe Pose and a simple classifier to detect good vs bad posture from a webcam stream. Includes tools to collect your own dataset and train a model.

• Real‑time webcam posture prediction
• Data collection script to label frames as good/bad
• Simple training script (Decision Tree) with saved model

• posture/ – Python package with reusable modules
  • geometry.py – simple Point and angle utils
  • collect.py – run_data_collector() to capture labeled frames
  • predict.py – run_live_predictor() for real‑time inference
  • train.py – train_model() to train and save a model
• data/ – dataset location (CSV), kept out of Git
• models/ – trained model artifacts, kept out of Git
• live_posture_predictor.py – Thin wrapper that runs the live predictor
• Python_data_collector.py – Thin wrapper that runs the collector
• train_model_v3.py – Thin wrapper that trains the model

1. Create and activate a virtual environment

python3 -m venv .venv
source .venv/bin/activate  # macOS/Linux
# On Windows: .venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt

3. Collect labeled data (optional, if you want to retrain)

python Python_data_collector.py

• Press g for good posture, b for bad posture, q to quit
• Data will append to data/posture_dataset.csv

4. Train the model (optional)

python train_model_v3.py

• Outputs models/posture_model_v3.pkl

5. Run the live predictor

python live_posture_predictor.py

• Press q to quit

• The included trainer uses a Decision Tree for simplicity. You can swap in any sklearn model.
• Features are the raw 3D pose landmark coordinates from MediaPipe; you can experiment with engineered angles.

• Python 3.9+
• Webcam access

• If the webcam doesn’t open, try changing the camera index in cv2.VideoCapture(0) to 1.
• If MediaPipe install fails on Apple Silicon, ensure pip is up to date and Python is from python.org or use conda.

This project is released under the MIT License. See LICENSE.