Deep Learning Based Camera Switching for Sports Broadcasting

Switching camera views when broadcasting sport games is essential for improved quality of viewing experience. Traditionally, expensive equipment and a broadcast director are employed to choose the optimal camera view during the game. In this study, we propose a novel deep learning based method to automatically switch between cameras based on importance of the scene detected in each view. Here, in order to show the validity of our approach, we chose to train our network for ice hockey, as the network needs to be retrained for each sport, using a dataset that corresponds to the specific game. Our method uses a YOLOv4 model that accurately detects the important objects for ice hockey, namely players, puck, net, goalie, and referee. Our novel camera switching view scheme uses the confidence values of the detected objects and temporal tracking to choose the most important camera view for that instance.