Master's Thesis : Deep learning for robust soccer pitch localization

Abstract

Sport field localization has many applications such as generation of statistics, autonomous camera systems or automated generation of highlights. Nowadays, not only large federations like FIFA are interested in the calibration of autonomous and manually steered cameras, but also smaller clubs want to have such a system for analyzing matches and making strategies. For all these applications, it is essential to have a robust pitch localization even in poor conditions, for example low lighting and inconspicuous lines.

EVS has developed an autonomous calibration system. This system is based on a neural network technique. However, during a Standard-Anderlecht soccer game filmed in poor conditions, the system repeatedly failed to localize the soccer pitch. The objective of this master thesis is to solve this problem in order to localize the soccer pitch even in such bad conditions.