Introduction

As with other sectors, AI is transforming the Sports Industry as well with applications in Cricket, Tennis and others along with Football. Computer Vision and Deep Learning algorithms are used to generate both predictive and descriptive analytics in various sports. In this post we will explore how we are leveraging computer vision to generate analytics from a football video.

Football Analytics Use Case

As a part of enhancing our capabilities in the AI & ML domain, Oneture Technologies is working on Football Analytics using Computer Vision use case. This project aims at generating 12 statistics and 3 predictions from any football match video by leveraging deep learning algorithms and computer vision technology.

Statistics:

1. Goals
2. Shots
3. Passes
4. Tackles
5. Interceptions
6. Saves
7. Fouls
8. Offsides
9. Corners
10. Freekicks
11. Yellow Cards
12. Possession Percentage

Predictions:

1. Expected Goals
2. Expected Possession Value
3. Expected Threats

This system aims at aiding as well as replacing the currently popular model of football analytics in which analysts manually tag players and other attributes during a football match

Leveraging Computer Vision to Generate Data from a Video

We are leveraging Computer Vision to generate raw data from a given football match video. For this, our ML Pipeline is divided in 7 components:

1. Object Detection - detecting players and ball in a video
2. Multi-object Tracking - tracking players and ball across frames
3. Player Identification - identifying a player across frames
4. Team Differentiation - differentiating between two teams in a video
5. Team Recognition - Recognizing team in a video based on jersey color and other parameters
6. Reference System & Homography Estimation - mapping video to 2-D pitch to generate raw data from video
7. Event Identification - Identifying various on-ball events like goals, passes, throws, shots, etc

Once raw data is generated from a video, it can be transformed to create tracking data and event data from which various statistics and predictions can be obtained.

Generating Predictions for any time slice

We are using various ML algorithms like XGBoost, Decision Tree and others to generate predictions like Expected Goals, Expected Possession Value, Expected Threats from the tracking and event data generated by the Computer Vision model.

Conclusion

Computer vision can be leveraged to generate data from any football videos which can be analyzed in order to provide statistics and predict outcomes. This process allows us to gain a better understanding of the game and its performance thereby helping players, coaches and other stakeholders.