AI-Gym

Real-time push-up coaching with MediaPipe 2D and Streamlit

AI-Gym is a Streamlit app for real-time push-up coaching using webcam-based pose detection and a MediaPipe 2D classification pipeline.

This repository contains:

• the live Streamlit app
• the active application pipeline in src/
• a retraining workflow in retraining/
• the original experimental notebooks and offline analysis in posture_checker_offline/

Final App Pipeline

The final application is explicitly 2D-only.

The live app uses:

• MediaPipe pose detection
• 2D landmark-based feature extraction
• repetition-level feature aggregation
• the trained model in data/models/pushup_2d.joblib

Setup

Clone the repository:

git clone https://github.com/Hayo87/AI-Gym.git
cd AI-Gym

Create and activate a conda environment:

conda create -n aigym python=3.10 -y
conda activate aigym

Install dependencies:

python -m pip install --upgrade pip
python -m pip install -r requirements.txt

Run the App

python -m streamlit run app.py

Then open:

http://localhost:8501

Allow webcam access when prompted.

Project Structure

• app.py Streamlit front-end for the live push-up coach.
• src/ Core application pipeline, pose processing, feedback logic, and shared utilities.
• retraining/ Scripts to regenerate features, retrain the 2D model, and run evaluation on filtered videos.
• data/models/ Saved models used by the app.
• data/processed/ Generated feature CSV files used for retraining.
• posture_checker_offline/ Experimental notebooks, offline analysis, and original project artifacts.

Retraining Workflow

The retraining workflow lives in retraining/ and reuses the live app pipeline so that training and inference stay aligned.

1. Generate Features From Raw Videos

This step:

• reads the raw videos from:
  • posture_checker_offline/data/raw/videos/correct
  • posture_checker_offline/data/raw/videos/incorrect
• runs them through the same app-side pose processing route
• creates a repetition-level feature CSV
• still writes a placeholder row for videos where no full repetition was detected

Run:

.\venv\Scripts\python.exe retraining\create_features.py

Output:

data/processed/mediapipe_2d_app_pipeline_features.csv

2. Train a New 2D Model

This step:

• reads the generated feature CSV
• removes rows without full repetition features
• trains an MLP on:
  • elbow_mean
  • elbow_std
  • body_mean
  • hip_mean
• saves the model for the live app

Run:

.\venv\Scripts\python.exe retraining\training.py --model-output data/models/pushup_2d.joblib

The training pipeline follows the notebook-style sklearn workflow:

• SimpleImputer(strategy="median")
• StandardScaler()
• MLPClassifier(hidden_layer_sizes=(64, 32), max_iter=1000, random_state=42, ...)

If a model already exists at the target path, the training script creates a backup with the _old suffix first.

3. Evaluate on Filtered Videos

The retraining test runner:

• reads data/processed/mediapipe_2d_app_pipeline_features.csv
• keeps only videos where at least one repetition was detected
• reruns those videos through the live pipeline
• checks whether correct videos stay correct and incorrect videos are detected as incorrect

Run:

.\venv\Scripts\python.exe -m pytest retraining\test_runner.py -s

Notes on Metrics

• Metrics reported by retraining/training.py are computed on the filtered repetition-level feature dataset.
• These metrics are useful for model comparison, but they are not identical to full end-to-end app performance on raw videos or live webcam input.
• End-to-end behavior also depends on detection quality, repetition detection, and the runtime pipeline.

Notes

• Python 3.10 is required for compatibility with streamlit-webrtc.
• Make sure your webcam is enabled.
• Stand far enough from the camera for reliable pose detection.
• The original model-development work is preserved in posture_checker_offline.