For a client building AI tools for soccer academies

The Problem

One of our clients wanted to build an AI tool to analyse soccer matches and automatically derive insights that can help players identify the areas in which they need to improve. The goal was to build an AI tool capable of extracting detailed analytics from soccer match videos, including player identification and tracking, field detection, and player statistics including the number of shots they played, the duration for which they had the ball with them and the number of successful passes and intercepts they were able to make.

Our Solution

We used a multi-stage approach for this:

1. Player and Ball Detection: For detecting players and the ball, we trained an object detection model based on YOLO architecture. For training, we annotated dozens of videos of real soccer matches played at various academies. After annotation, these videos were used to train the model.

2. Player and Ball Tracking: We used DeepSort to track players and the ball through the match.

3. Player Re-identification: Since the camera only focuses on a part of the field at any time, so players often appear and disappear from the camera view. This means that we need a way of re-identifying a player when they re-appear on the camera. Because the videos we process are low quality, so the players' faces are not always very clearly visible, meaning that we cannot use facial recognition for this task. Instead we use a clustering technique based on different features including colour histograms to detect when the player reappears. 

4. Team Detection: We also use a separate ML model to detect which team a player belongs to by looking at the colour the player's shirt.

5. Event Detection: Once we know the location of the ball and which player has the ball at any time, then we can easily detect events such as passes, intercepts and dribbles. For example, a pass is when the ball moves from one player to another player of the same team.

Results

Our soccer analytics tool achieved:

- Accurate player and ball tracking for in-depth performance analysis.

- Reliable player re-identification, facilitating team association during matches.

- Comprehensive event detection, providing insights into key match moments for tactical improvements.