Master Thesis presentation by Emil Holmberg and Daniel Illipe

Examiner: Lennart Svensson

Abstract:
Sports analytics is a growing industry which relies on data gathered from games. In this thesis, we construct a pipeline for extracting 2D and 3D data about all players' positions and the ball's position in football games. To this end, we perform camera calibration and use triangulation of 2D pose estimates from multiple angles to extract 3D poses. The first step of the pose estimation is object tracking which is performed using a YOLOX model finetuned for this specific problem setting. The 2D pose estimation is then done using ViTPose which is a simple and robust model for 2D pose estimation that uses vision transformers. As input ViTPose uses the results of the object tracker by using a player detection for each 2D pose predicted. A method for the more challenging problem of tracking the ball position in 3D is presented through a combination of triangulation, homography and logic. As a byproduct of extracting this data, we construct animations of game sequences that could be used as a tool in football analytics. Additional side studies are performed on classifying the players into their corresponding teams with Protonet, a few-shot learning model, and automatically calibrating cameras, mainly considering template matching using a similarity measure.

Welcome!
Emil, Danie and Lennart