Résumé: In the tile-based panoramic video streaming, the Field of View (FOV) is composed of multiple real-time synchronized visible video tiles. The common panoramic video transmission control methods use the FOV prediction and redundant tile transmission to address the issues of network delay and fast viewport switching. However, these methods rely heavily on the FOV prediction accuracy and do not fully consider the transmission efficiency, which is measured by the ratio of data used for FOV to the total transmitted data. Moreover, the existing learning-based methods directly consider the ever-changing factors such as network bandwidth and viewport position in the learning process, resulting in the poor stability of the transmission control. In this paper, we propose a Deep Reinforcement Learning (DRL)-based transmission control method for the tile-based panoramic video streaming, and the objective is to optimize the transmission efficiency on the basis of the guaranteed Quality of Experience (QoE). Firstly, we define the panoramic video transmission control process as the maximization of transmission efficiency on the basis of constraining multiple QoE metrics in the preset acceptable ranges. Secondly, we design a two-stage transmission control decision-making mechanism to improve the stability of transmission process, which includes intermediate decision-making stage and final decision-making stage. During the intermediate decision-making stage, the newly defined aggregated transmission decision variables are learned by using the Rainbow Deep Q Network. In this online learning process, we only consider the QoE and transmission efficiency, and avoid directly involving the ever-changing environment factors. During the final decision-making stage, the bitrate and buffer size of each video tile are determined according to the network bandwidth and viewport under the guidance of the intermediate decisions. Finally, the experiments conducted with the actual network bandwidth and viewport track show that our method performs better in the long-term transmission efficiency than other methods.

Langue: Anglais