Identification of Team Tactical Formations and Player Positions in Association Football

This thesis proposes a method for identifying team tactical formations and player positions in association football. Chapter 1 introduces formations as concise and practical abstractions that summarize a team’s spatial arrangement, commonly reported using three to five digits that represent the number of outfield players, from defense to attack, in a symmetric manner. These formations, such as the Pyramid, 3-2-2-3, and 4-4-2, have been present since the early days, as seen in football history books, and serve as a reference for players when they get distracted or to create desired scenarios like numerical superiority. While formations are important for opposition analysis, the media’s starting formations are not a reliable source because they are just pre-match predictions and do not reflect in-game dynamics. In Chapter 2, we provide a comprehensive review of previous formation identification studies and categorize their main steps into preprocessing tracking data and following either a team- or position-level approach. Preprocessing includes segmenting match time and normalizing player locations. At the team level, the formation is reported for the entire team at once, while at the position level, individual player positions are inferred and then the formation is subsequently reported as a collection of these positions. Both levels first choose a data representation and later perform template matching or clustering. This chapter also discusses key assumptions, implications, and potential improvements, including alternative evaluation methods in the absence of ground-truth labels. In Chapter 3, we introduce shape graphs as a fundamental structure to identify player tactical positions and derive team formations. Starting from the Delaunay Triangulation, the dual of the Voronoi diagram (a simple space control model), we first quantify their edge stability and later iteratively remove edges sensitive to small player location perturbations. The resulting subgraph, shape graph, offers a stable representation (almost) preserving nearest teammates and convex face properties, while the edge changes of the shape graph and Delaunay graph occur at different times across a match, as empirically shown. In Chapter 4, we infer each player’s tactical position at the highest available data resolution by coloring the shape graph’s vertices along the vertical and horizontal axes, using face centers as references. A comparison of our assigned labels with publicly available operator tags reveals a 70% agreement and provides plausible explanations for the remaining differences. In Chapter 5, we derive formations from the inferred positions at three levels: 1) overall setups with position maps, 2) match segments using position movement diagrams, and 3) instantaneous by applying our proposed exact and relaxed matching methods on a comprehensive set of formation templates. Although exact matching explains only a small fraction of match time and even relaxation (horizontal, corner, and vertical) increases this to around 10%. Our proposals reduce noise and better align with analysts’ video-based workflows, unlike the nearest neighbor and grouping vertical coordinates methods. Finally, we introduce codes as a generalization of these formation templates. In Chapter 6, we explain match/opposition analysis and player positional scouting applications of shape graphs. We are particularly proud of position maps, position plots, and position movement diagrams. Position Maps profile players’ position occupancy across matches, supporting the contextualization of performance for scouting. Position plots are time series visualizations of players’ inferred positions, taking into account relevant contexts such as goals, bookings, stoppages, and substitutions, making them relevant for match/opposition analysis. Finally, the position movement diagrams offer a spatial representation of position plots across match segments.