Automatic detection of changes in the real tactical formation

The real tactical lineup (RTL) is an analysis option that can be used to illustrate the real tactical action on the field. In contrast to the tactical formation given by the trainer (see Figure 1, left), the
RTL (right) reflects the actual positions of the players. To determine the
RTL, representative positions are determined for the players that represent the actual stays over a time interval, e.g. a half-time or the entire game. Usually, all player positions in the corresponding period are averaged. The
RTL is thus composed of the mean player positions.

Figure 1: Comparison of a tactical formation (left) to a real tactical lineup (right). The latter shows the actual player positions on the field. 

Since there may be temporary changes in the formation of a team during a game for various reasons, for instance, two players change their roles for a few minutes, there will be a problem when analyzing the teams solely based on the averaged player positions. In the (slightly exaggerated)
scenario shown in Figure 2, the red highlighted player owns the position of a central-defender during the first half of a half. In the second half he becomes the second attacker. When calculating an averaged lineup based on this scenario, the player’s will become a central midfielder for the entire half. However, this does not correspond to reality, as he has never held this position at any time.

Figure 2: Temporary changes in the formation of a team may lead to problems using a RTL-aided game analysis.

One solution to this problem would be to create a RTL for the time intervals between the mentioned changes and to analyze them afterwards separately from each other. However, a problem is that the start and end times (interval limits) of the changes are usually not known or would have to be determined by a prior analysis of the player’s positions. One approach, which integrates an automatic determination of the interval limits for each player individually in the calculation of a RTL, consists of using unsupervised machine learning methods. In this approach, a clustering method is applied to each player’s trajectory. In this way, the different hotspots of a player can be identified. Subsequently, these hotspots will be temporally ordered, if they have a significant length of stay. The resulting sequence of hotspots can be visualized like a RTL. The mean positions of the players are replaced by their hotspot sequences. In this way, the resulting graphics (see Figure 3) also shows the players’ roles over time. Determining a position that a player has never owned in reality is thus bypassed.

Figure 3: The automatically generated history of different real tactical lineups reveals temporary changes in the formation. The sequence of player hotspots is visualized by their markers linked with arrows.