Football: Analysing games with artificial intelligence
Artificial intelligence and machine learning can be very helpful when it comes to recognising tactical patterns in the analysis of football matches. This is shown by a paper that has now been awarded 1st prize at the MIT Sloan Sports Analytics Conference in Cambridge (Massachusetts). Its authors include computer scientists Prof. Dr. Ulf Brefeld and Dennis Faßmeyer from Leuphana University Lüneburg.
Video analysis departments are an established part of professional football. In order to gain meaningful insights into a team’s tactics, experts usually look for tactical patterns manually when evaluating video footage and comment on them. This is a time-consuming and repetitive process. The scientists‘ idea was that a lot of time could be saved by at least partially automating this process with the help of artificial intelligence.
If you want to recognise tactical patterns and behaviours on the football field based on tracking data, you usually need a lot of already classified patterns that have emerged from video analyses. Together with two colleagues from the German Football Association (DFB) and Hertha BSC Berlin, Brefeld and Faßmeyer have developed a method that requires only a very few previously hand-selected scenes for the analysis and can also use unclassified situations. Their so-called „autoencoder“ can recognise tactical patterns at team, group and player level with the help of a graph neural network and automatically find relevant scenes of a match.
For their work, they used the example of the overlap run during matches of the German national football team. This is an established tactical pattern in football involving two players: The ball carrier dribbles the ball (usually in the opponent’s half close to the sideline), a teammate runs past him at high speed and creates a passing option to the opponent’s goal. Such runs are often difficult to defend, so it is an advantage in tactical match preparation to know which opposing players are typically involved in such a pattern.
With the new method, corresponding information can be created automatically for each upcoming opponent. An overview then shows which players overlap and who is overlapped the most. In this way, one also learns about the preferences of pairings that make overlapping runs together.
In their future work, the scientists want to look at recognising a variety of other patterns, such as shots on goal, counter-attacks or the defence of corner kicks, and transfer the approach to other team sports such as basketball.