Fabian Hirn

Advisors

Rebecca Lennartz (M. Sc.),  Alexander Weiß (M. Sc.), Prof. Dr. Anne Koelewijn, Prof. Dr. Björn Eskofier

Duration

07 / 2024 – 09/ 2024

Abstract

Improved data collection methods and increased computational power allow for superior performance analysis in soccer [1]. A well-known, but much discussed metric in this field, Expected Goals (xG), is based on a machine learning classification algorithm that calculates the probability of a shot being converted [2]. The most basic models use the distance to the goal, the angle to the goal, and how much pressure the shooting player experiences as features [3, 4], whereas more advanced approaches also consider the opponent’s position through tracking data [5]. While early models mainly used logistic regression as the classification algorithm [3, 6], recent work focused on Gradient Boosting methods [5].

One major downside of the xG-model is that the resulting value only represents the average of similar shooting situations in the dataset [7]. Therefore, a tendency exists for better-regarded goal scorers to overperform their xG-value [8], which suggests that unknown factors not considered in the current models are relevant for predicting goal probabilities. It is assumed that the deviation of certain players from their xG-value can be explained by their finishing skill. Hewitt et al. [8] proposed to calculate position-adjusted models for each main positional subgroup (Defenders, Midfielders, and Forwards) to address this problem. But this approach was still not able to explain the individual deviations between the players in each subgroup.

The finishing skill can be analyzed by a player’s shooting motion. Research demonstrated a correlation between certain biomechanical features of the shooting motion, such as the knee angle or degree of flexion in the support leg during an instep soccer kick [9, 10], and accuracy. For example, it is suggested that an approach angle of 45° to the ball is optimal for both the instep and inside scoring techniques [10, 11, 12]. One method to record such biomechanical features is Pose Estimation [13]. Pinheiro et al. [14] used it combined with Notational Analysis to determine whether biomechanical features, such as the knee angle of the support leg, correlate with the penalty taker’s and goalkeeper’s strategies in a shoot-out. The usage of such features as input to xG-models is not known to the author.

Thus, the aim of this thesis is to build a machine learning-based classification model for Expected Goals using biomechanical features derived from Pose Estimation. Consequently, one research question of this thesis is whether it is possible to extract the selected biomechanical features from Pose Estimations of shooting motions. Additionally, whether these features being used in an xG-model combined with the established features of previous models improve the accuracy. Furthermore, it is investigated whether the usage of these features leads to a model for player-specific shooting accuracy.

References

[1] D. Memmert, R. Rein, “Match Analysis, Big Data and Tactics: Current Trends in Elite Soccer,” German Journal of Sports Medicine, vol. 69, no. 3, pp. 65-72, 2018.
[2] S. Green, “statsperform.com,” StatsPerform, 12 4 2012. [Online]. Available: https://www.statsperform.com/resource/assessing-the-performance-of-premier-league-goalscorers/. [Accessed 15 4 2024].
[3] R. Pollard, J. Ensum and S. Taylor, “Applications of logistic regression to shots at goal in association football: calculation of shot probabilities, quantification of factors and player/team,” Journal of Sports Sciences, vol. 22, no. 6, p. 504, 2004.
[4] “bundesliga.com,” DFL Deutsche Fußball Liga GmbH, 1 7 2021. [Online]. Available: https://www.bundesliga.com/de/bundesliga/news/expected-goals-xgoals-torwahrscheinlichkeit-analyse-statistik-aws-amazon-13252. [Accessed 15 4 2024].
[5] D. Vatvani, “Upgrading Expected Goals,” Statsbomb, 16 5 2022. [Online]. Available: https://statsbomb.com/articles/soccer/upgrading-expected-goals/. [Accessed 15 4 2024].
[6] M. Caley, “Premier League Projections and New Expected Goals,” SB Nation, 19 10 2015. [Online]. Available: https://cartilagefreecaptain.sbnation.com/2015/10/19/9295905/premier-league-projections-and-new-expected-goals. [Accessed 17 4 2024].
[7] J. Davis and P. Robberechts, “Expected Metrics as a Measure of Skill: Reflections on Finishing in Soccer,” in 10th Workshop on Machine Learning and Data Mining for Sports Analytics, Turin, 2023.
[8] J. H. Hewitt and O. Karakuş, “A machine learning approach for player and position adjusted expected goals in football (soccer),” Franklin Open, vol. 4, 2023.
[9] A. Arguz, A. Guebli, N. Erkmen, S. Aktaş, M. Reguieg and Y. Er, “Biomechanical analysis of accuracy penalties-kicking performance for Turkish Soccer players: Group-based analysis without goalkeeper,” Physical Education of Students, vol. 25, no. 3, pp. 189-196, 2021.
[10] E. Kellis and A. Katis, “Biomechanical characteristics and determinats of instep soccer kick,” Journal of Sports Science and Medicine, vol. 6, pp. 154-165, 2007.
[11] J. Scurr and B. Hall, “The effects of approach angle on penalty kicking accuracy and kick kinematics with recreational soccer players,” Journal of Sports Science and Medicine, no. 8, pp. 230-243, 2009.
[12] B. Mihailović, L. Lilić, R. D’Onofrio, T. Koliopoulos, M. Pal and G. S. Iacob, “Biomechanics of Kicks in Football: A Review,” Italian Journal of Sports Rehabilitation and Posturology, vol. 10, no. 26, pp. 2779-2791, 2023.
[13] M. Sypetkowski, G. Kurzejamski and G. Sarwas, “Football Players Pose Estimation,” in Image Processing and Communications Challenges 10, Springer, 2019, pp. 63-70.
[14] G. d. S. Pinheiro, X. Jin, V. T. Da Costa and M. Lames, “Body Pose Estimation Integrated With Notational Analysis: A New Approach to Analyze Penalty Kicks Strategy in Elite Football.,” Frontiers in Sports and Active Living, vol. 4, no. 818556, 2022.