Center of attention prediction for automatic video framing in teamsport events

Keemotion is a spin-off from Université Catholique de Louvain proposing automated sport production. One of the key functionality of their product is the automated video framing of the filmed game: based on the sequences recorded by a fixed camera above the court, an algorithm is responsible for planning where the camera should look. The current solution implemented by Keemotion is based on a set of fixed rules and heuristics and, although it is rather effective for most situations, it showed some weaknesses in particular tricky game phases. In this work, we propose a method to predict an important information for the framing: the Center of Attention, which could be used to improve current Keemotion's system or implement a new framing strategy. The problem will thus be formulated as a supervised learning task: using a set of framing examples manually annotated and the associated visual characteristics extracted automatically from the scene, we train a regression function which will be able to predict the Center of Attention in new video sequences.