FLUS-FoV: User Scheduling for Federated Learning-Based 360-Degree Video FoV Prediction

The accurate prediction of the Field of View (FoV) is instrumental in optimizing the streaming performance of 360-degree videos, considering that users typically focus on specific portions of the video (denoted as FoV)and periodically switch between them. However, when privacy concerns restrict the sharing of user view data, the inability to utilize this data for training network models significantly hampers the precision of FoV prediction. To tackle this challenge, this paper introduces an innovative federated learning-based FoV prediction algorithm. This algorithm harnesses a combination of 360-degree video saliency and user view trajectories to forecast FoV with precision. Furthermore, it presents a novel algorithm for quantifying the information value with communication resource allocation to optimize this prediction process. Through experimental assessments, our findings demonstrate that the proposed scheduling algorithm outperforms traditional methods, and the distributed FoV prediction accuracy closely approximates centralized approaches.