- Track the ball
- Detect court lines
- Detect the players
To track the ball we used TrackNet - deep learning network for tracking high-speed objects. For players detection yolov3 was used.
Example using sample videos
| Input | Output |
|---|---|
.gif) |
 |
.gif) |
.gif) |
 |
 |
This project requires compatible GPU to install tensorflow, you can run it on your local machine in case you have one or use Google Colaboratory with Runtime Type changed to GPU.
Input videos should be rallies of the game and should not contain any commercials, breaks or spectators.
- Clone this repository
- Download yolov3 weights (237 MB) from here and add it to your Yolov3 folder.
- Install the requirements using pip
- Run the following command in the command line
- If you are using Google Colab upload all the files to Google Drive
-
Create a Google Colaboratory Notebook in the same directory as
predict_video.py, change Runtime Type to GPU and connect it to Google drive -
Change the working directory to the one where the Colab Notebook and
predict_video.pyare. In my case, - Install only 2 requirements, because Colab already has the rest
-
Inside the notebook run
predict_video.py
git clone https://github.com/ArtLabss/tennis-tracking
pip install -r requirements.txtpython predict_video.py --input_video_path=VideoInput/video_input3.mp4 --output_video_path=VideoOutput/video_output.mp4 --minimap=0 --bounce=0from google.colab import drive
drive.mount('/content/drive')import os
os.chdir('drive/MyDrive/Colab Notebooks/tennis-tracking')!pip install filterpy sktime !python3 predict_video.py --input_video_path=VideoInput/video_input3.mp4 --output_video_path=VideoOutput/video_output.mp4 --minimap=0 --bounce=0
After the compilation is completed, a new video will be created in VideoOutput folder if --minimap was set 0, if --minimap=1 three videos will be created: video of the game, video of minimap and a combined video of both
P.S. If you stumble upon an error or have any questions feel free to open a new Issue
- Court line detection improved
- Player detection improved
- The algorithm now works practically with any court colors
- Faster algorithm
- Dynamic Mini-Map with players and ball added, to activate use argument
--minimap
--minimap=0 |
--minimap=1 |
|---|---|
 |
 |
- By specifiying
--bounce=1bounce points can be detected and displayed
To predict bounce points machine learning library for time series sktime was used. Specifically, TimeSeriesForestClassifier was trained on 3 variables: x, y coordinates of the ball and V for velocity (V2-V1/t2-t1).
Data for training the model - df.csv.
The model predicts true negatives (not bounce) with accuracy of 98% and true positives (bounce) with 83%. As a result, one or two bounces might be overlooked by the model.
Improve line detection of the court and remove overlapping linesAlgorithm fails to detect players when the court colors aren't similar to the sample videoDon't detect the ballboys/ballgirlsDon't contour the bannersFind the coordinates of the ball touching the court and display them- Code Optimization
Dynamic court mini-map with players and the ball
- Slow algorithms (to process 15 seconds video (6.1 Mb) it takes
28 minutes16 minutes) - Algorithm works only on official match videos