Artificial Intelligence for the automation of robotic cameras in live sports

摘要

The demand to cover lower tier sports games is rising rapidly. Unfortunately, it is economically not viable to deploy the traditional production methods, crews and equipment that we know for higher tier sports events. Production approaches requiring much less people are needed. For the last few years, high quality single man operated equipment combining replay video mixing, graphical insertions, … are being introduced. However, one still needs cameramen to operate the different cameras around the sports field. — In this paper we present the concept of a fully automated system creating multiple camera feeds via a series of robotic cameras. These robotic cameras are not human operated but piloted by an Artificial Intelligence based system. This system analyses the complete sports scene for relevant items to capture and uses that information to individually steer the different robotic cameras to generate the desired different camera views. More concretely, the AI backed system points, in real-time, the different robotic cameras to the relevant parts of the game and applies the appropriate zoom factor corresponding to the desired camera views. We will show first results of a prototype system applied to a soccer game. — To further push the limits of automating the production, the automatic camera operation is complemented with an automatic selection of the most interesting camera angle out of the available camera feeds. This automatic direction can either be used to create a fully automatic system or to assist the single operator in charge of a complete multi-camera production. First experiments indicate that the quality of the camera selection comes close to the judgement of real directors. — To achieve human operator quality, both for the camera selection and for the automatic steering of robotic cameras, we designed an Artificial Intelligence based approach, combining several techniques, including a combination of various machine learning components to mimic the behavior of human operators.