Author(s): PEI, Y., CHAOYI, G., JIAMING, N., VARUNA, S., Institution: LOUGHBOROUGH UNIVERSITY - LONDON, Country: UNITED KINGDOM, Abstract-ID: 849

INTRODUCTION:
Sports performance analysis has been developed for decades, and essential metrics like passing rates and tackles have been used to evaluate players, while the context of these metrics is ignored. The key difference between an average player and a star player is how they handle pressure, make correct decisions and keep the same elite performance. Therefore, this study aims to quantify the pressure context for football players and analyse their performance under different pressure levels.
METHODS:
Tracking and event datasets with broadcasting videos of 10 different Premier League matches have been used in this research. Tracking data records the x and y location of each player and the ball on the pitch at 25 Hz per second and the event dataset contains the on-ball events including passes, shots and tackles during the match. A Graph Neural Network (GNN) based sequential model is applied to predict whether the possession team is going to lose control of the ball under the opponent’s pressure. A higher likelihood of losing control of the ball equals a higher level of pressure, and players’ performance under different levels of pressure is then evaluated. Deep learning, GNN-based sequential model, 3D Human Body Estimation and Player Pressure Map (PPM) features were used to leverage the dataset and predict the outcome of football footage.

RESULTS:
The purpose of this model is to predict whether the possession team will lose the ball control or not within seconds. The result is compared between various models and features. For the basic GNN model with only a tracking dataset, the prediction accuracy is 55.8%. By adding the 2D PPM features we generated, the accuracy increased to 75.2% and when the PPM features turned into 3D, the accuracy raised to 78.7%. By upgrading the model to a sequential GNN, the prediction accuracy with 3D PPM features reaches 81.2% which is the highest among all the experiments—both PPM features and time-sequential model help to improve the model accuracy massively.

CONCLUSION:
The contribution of this pilot work was to quantify the pressure level based on the football context and time-sequential dataset with 3D human body pose. The work is aimed to represent the 3D football context and predict the pressure level of football matches. The pressure metric and context performance analysis are based on the prediction. In summary, PPM features and time-sequential dataset help to improve the model accuracy and lead the metric step forward.