Towards the automatic selection of cameras for the production of live television shows

Abstract

In order to produce entertaining broadcast shows, most soccer games are captured by multiple cameras. During the game, a technical director chooses at any time which camera to broadcast. This master’s thesis introduces a framework that can be used to assist or replace the technical director in the task of camera selection. This framework uses computer vision components in order to extract information about the context of the game and the type of image captured by each camera. Machine learning is then used to rank the cameras according to their suitability for broadcast. Some heuristic rules are also inserted in order to guarantee some smoothness in the camera ranking. Compared to the previous approaches for the automatic selection of cameras, our framework uses information about the events happening on the field to adapt the camera selection strategy. Several parameters can be modified in order to produce videos according to user’s preferences. The quality of the videos produced by the system was evaluated with a user study, and the results were very promising. On video clips of one minute, the videos of the technical director were preferred to the videos of our system only half of the time. Moreover, on a scale of 1 to 5, the average score of the technical director’s videos was only 0.4 higher than that of our system’s videos.