This repository applies computer vision techniques to extract valuable insights from a padel game recording like:

• Position and velocity of each player;
• Position and velocity of the ball;
• 2D game projection;
• Heatmaps;
• Ball velocity associated with distinct strokes;
• Player error rate.

To do so, several computer vision models where trained in order to:

1. Track the position of each individual players;
2. Players pose estimation with 13 degrees of freedom;
3. Players pose classification (e.g. backhand/forehand volley, bandeja, topspin smash, etc);
4. Predict ball hits.

The goal of this project is to provide precise and robust analytics using only a padel game recording. This implementation can be used to:

1. Upgrade live broadcasts providing interesting data to be shared with the audience or to be stored in a database for future analysis;
2. Generate precious insights to be used by padel coachs or players to enhance their path of continuous improvement.

conda create -n python=3.12 padel_analytics pip
conda activate padel_analytics
pip install -r requirements.txt

The current model weights used are available here https://drive.google.com/drive/folders/1joO7w1Am7B418SIqGBq90YipQl81FMzh?usp=drive_link. Configure the config.py file with your own model checkpoints paths.

At the root of this repo, edit the file config.py accordingly and run:

python main.py

Using the default batch sizes one will need to have at least 8GB of VRAM. Reduce batch sizes editing the config.py file according to your needs.

Currently this implementation assumes a fixed camera setup. As a result, a UI for selecting court keypoints will pop up asking you to select 12 unique court keypoints that are further used for homographic computations. A video describing the keypoints selection is available at ./examples/videos/select_keypoints.mp4. Please refer to main.py lines 24-38 where a diagram showcasing keypoints numeration is drawn.

I am currently looking for collaborations to uplift this project to new heights. If you are interested feel free to e-mail me at [email protected].