Towards automatically classifying football formations for video analysis

Abstract

Formations form the tactical basis of football analytics. Formations influence how aggressive a team plays, how they create chances and how they defend. Because of the increased availability of positional data, there has been a rise in the research towards automatically recognizing football formations. However, most of the previous research has not been clear on the effectiveness of their algorithm and has not shown how these algorithms can be used by domain experts. This study has three aims. First, it reports on the accuracy of three previously proposed algorithms by Shaw and Glickman (2019), Muller-Büdack et al. (2019) and Narizuka and Yamazaki (2019). to show their effectiveness. Secondly, it proposes several improvements for these algorithms. Lastly, it shows how these algorithms can be incorporated into the workflow of domain experts. In this research, the existing algorithms of Müller-Budack et al. (2019) and Shaw and Glickman (2019) perform between 13% and 28% accurate respectively. respectively on a difficult dataset based on Wyscout reports. I propose several improvements which increase the accuracy of the best performing algorithm to 45% and which showed a promising 78% in a case study. Finally, I examined how these algorithms could be used by domain experts. First, a method is proposed to convert the formation analysis into an XML, which can be imported into the video tools used by the domain experts. Secondly, I show how these algorithms can be used for data analysis.