Tracking de jugadors en imatges amb transformers

Abstract

[en] Data is becoming increasingly important across the board, and so is within sports disciplines. Particularly, in football. For instance, the position of the different players on the pitch is in itself data which has been proved useful for applications such as injury prevention, player scouting and so on. This type of data, designated “tracking data”, is obtained at the time of this study using GPS technology on a professional level. Since local leagues and football institutions are responsible for the management of this data, it can prove to be difficult for outside actors to obtain access to it. In reality, this means this data ends up being accessible to high income and top league clubs and institutions only. For this very reason, the need to find alternative ways to generate and obtain such data arises. This project focuses its scope on using computer vision as an alternative to the previously stated. The aim of this work therefore, is to beforehand acquire a theoretical view and understanding in the branch of machine learning and convolutional neural networks and their application to detect people in football videos. In particular we will explore the use of Transformers, an architecture of convolutional neural networks that appeared very recently and involved a paradigm shift in state of the art models to process the source material and generate the raw data. With a specialized dataset of exclusively football images, we have been able to train a DETR model and compare its results with other existent models as a reference. With the results in hand, we have explored ways of improve such models. We have obtained a trained model that successfully manages to detect the players on a football match with the caveat of it being dependent on the source material’s quality. While it does succeed in most regular game play, it struggles in situations where the source material presents occlusions and accumulations of players in the image (for example during a corner kick, where many players can accumulate near the goal). We have been able to even slightly improve the results of the trained model by managing double detections, but the size of our dataset has proved to be a constraint in this direction. Finally, we have discussed some possible future lines of improvement to achieve better results, such as increasing our dataset or using a wider range of frames to reduce the margin of error when players are occluded.