Modelling motion tracking data in elite soccer to classify and quantify collision intensity

Motion tracking data is collected across an increasing number of sports. Some sports can be quite physical and can involve frequent collisions during both training sessions and matches. Depending on the motion tracking devices available for each sport, particularly in matches, it may limit the ability to accurately measure the severity of these collisions. Accelerometers can be used to measure the impact of collisions but they are not used in most sports. The development of a Collision Severity Index (CSI) using motion tracking data would allow analysts to measure collision severity without the use of accelerometers, which is currently absent from any literature. Therefore, the primary aim of this thesis is to present a novel approach to quantify collision severity in soccer by developing a CSI (Chapter 6) using Principal Component Analysis (PCA). In-game motion tracking data is used, which was collected from the English Premier League (EPL), the top soccer league in England. Loading scores can then be developed using the CSI, which can then be used to monitor the players during matches. The loading scores developed as part of this thesis were solely based on the accumulation of the collision severity scores for each player within each game. The loading scores could provide useful information for when players should be rested, which may lead to a reduction in the overall injury incidence in soccer and enhance the overall welfare for players in the sport. Using video footage, collision-related events were manually labelled across 20 matches in the EPL and all of the studies which involved using motion tracking data were based on 9 of these matches as this was sufficient for the studies. In the first study, all of the collision-related events in a typical soccer game were labelled in order to develop a Soccer Match Collision Dataset (SMCD) (Chapter 3), which proved to be essential for the remainder of the studies in the thesis, mainly for validation purposes. The next two studies involved feature extraction and event detection (for collision-based events) (Chapter 4) and developing tackle and collision classification models (Chapter 5). The aim of the feature extraction and event detection study (Chapter 4) was to firstly develop features that could be used for later studies in the thesis, including the development of classification models and the CSI. Secondly, the aim was to develop event detection algorithms for collision-based events in soccer matches using motion tracking data which included tackles, headers, set-pieces (corners, goal-kicks, free-kicks and throw-ins) and collision detection algorithms. Headers, corners, goal-kicks, free-kicks and throw-ins could be detected with the highest level of accuracy. However, there were many false positives for the tackles and collisions. Developing event detection algorithms for tackles and collisions proved to be the most challenging. Given the diverse nature of tackles and collisions in the sport and the limitations with the data (one point per player, no player elevation, etc.), it made it extremely difficult to detect tackles and collisions. There was also a lot of borderline cases for both tackles and collisions which added to the complexity of the study. The aim of the tackle and collision classification study (Chapter 5) was to develop classification models for the various tackle and collision types in soccer. Features were developed before and after the midpoints of the events in the hope to capture the motion of the players (and ball) in order to successfully develop the classification models. The study played a key part for a streamlined approach for the development of Collision-Specific Severity Indices (CSSI) as well as a general Collision Severity Index (CSI) for the collisions in soccer. Accurate tackle classification models could be developed. The accuracy was not high for the collision classification models due to the fact that a lot of the sub-classes were included in the body to body contact class, making it harder to make accurate predictions. There were not many observations for some of the collision classes also i.e., push and pull. Including all of the tackles and collisions that should really be classified in an 'other' class, also added to the complexity of the study. The fourth study (Chapter 6) was the main goal of the thesis; the development of a Collision Severity Index (CSI) in soccer using in-game motion tracking data. Many features were developed around the point of contact of the collisions and Principal Component Analysis (PCA) was used on different combinations of features (and variations of frames) in order to develop a CSI. A general CSI was developed as well as Collision-Specific Severity Indices (CSSI) for different collision types in soccer. The accumulation of all of the collision severity scores from the collisions could then be incorporated into a loading score for each player. The loading scores could potentially be used by coaching staff to guide them in the decision-making process on when to rest (or substitute) players during matches as well as whether to start or rest players before a match has commenced. A dashboard was then developed using R Shiny (Chapter 7), where a collision-hurt map (CHM) was the main feature. Such a tool can be used by analysts, researchers and the medical team in order to inform post game rehab, identify players at increased risk of injury and ultimately improve player safety and welfare. The general CSI (and CSSI) were proven to be very effective when the severity quantifications were analysed against the collision severity levels (heavy contact, medium contact, light contact and barely contact) using box plots (and violin plots) (Chapter 8). The results of this thesis have shown that an accurate Collision Severity Index (CSI) (and Collision-Specific Severity Indices (CSSI)) can be developed in soccer using in-game motion tracking data. These indices coupled with loading scores could potentially be used to guide coaching staff on when players should be rested, which could minimise the risk of injuries and improve the player welfare in the sport. Future studies are warranted in order to use this novel approach to quantify collision severity to aid with any collision-based tools which are currently absent from any literature using motion tracking data.

Funder

College of Science, University of Galway

Publisher

University of Galway

Rights

Attribution-NonCommercial-NoDerivatives 4.0 International

cbn

Collections

University of Galway Theses (PhD Theses)