3D-Virtual-Training-Coach

This repository contains code for the methods described in the following paper.
The code extract 3D skeletons through MeTRAbs from homemade videos frames to establish which parts of a fitness exercise are not correctly executed.

System Requirements

For GPU enviroment:

• CUDA 11.0 or greater versions;
• TensorFlow-gpu 2.5.0 or greater versions.

For CPU enviroment:

• TensorFlow 2.5.0

To generate 3D skeletons is necessary to download some pre-trained models: before running the code, you have to launch the method download_models().

Running the Code

First of all you have to generate 3D skeletons of fitness exercise frames running the generate_dataset( frames_path ) method: make sure that the Frames directory contains a sub-directory of the specific exercise frames, i.e Frames/arm-clap_1/frame.jpg.
Then normalize your 3D skeletons through normalize_dataset() method.
Finally choose an exercise to analyze with one of the metrics defined in the paper launching one of the following methods:

• identify_euclidean_errors();
• identify_angles_errors();
• identify_combined_errors().

Running Sample

You can try it in action using the try_me() method: it will guide you to run the code in a user friendly way with a minimal graphic interface.

How to use:
1. Select the directory containing the frame exercises directories;
2. Choose an exercise from the list;
3. Click on Next!

4. Select the metrics;
5. Set the thresholds;
6. Decide if you want to visualize the images of the errors;
7. Click on Analyze!