Viewport-Adaptive 360-Degree Video Streaming

With the rapid development of virtual reality technology and the gradual deployment of 5G networks,360-degree video will occupy an increasingly important position in network video services.In order to provide viewers with a real sense of experience in a virtual environment,360-degree videos usually require ultra-high resolution and frame rate to cover the full range of views of the video.These two prerequisites affect the bandwidth and bandwidth required for 360-degree video transmission.Server storage capacity presents new challenges.In order to provide users with an immersive experience while reducing bandwidth occupation and server storage capacity,this paper has carried out a research on the field of view adaptive360-degree video transmission method.On the one hand,by assigning different tiles to different tiles in the user's current field of view.The bit rate further reduces the wireless network bandwidth and server storage capacity occupied during 360-degree video transmission.On the other hand,a field of view prediction method based on deep learning is proposed,which uses the user's historical head movement to predict the user's field of view position in the future,which further improves the accuracy of field of view prediction.The main work of this paper is as follows:(1)An optimized field of view adaptive 360-degree video transmission method is proposed.This method is based on high-efficiency video coding HEVC Tiling,Motion-Constrained Tile Sets(MCTS),and MPEG-DASH spatial relationship description SRD.First,the visual importance level of different tiles in the f ield of view is proposed and the relationship is described.The adaptive MPD composition method is then encoded into three different bitrate versions,and then HEVC Tiling is used to evenly divide the video of each bitrate version into 3×4 or 3×5 or 3×6 ti les,Finally,different bit rates are assigned to different tiles in the user's field of view.The final experimental results show that this method is superior to the existing representative methods in terms of bit rate saving and storage capacity saving r ate while ensuring the viewing effect of users.(2)A gated recurrent unit(GRU)field of view prediction method embedded with a SENet(Sequeeze-and-Excitation Network)structure is proposed.The method first uses the GRU channel to extract the features in the field of view,and then uses the SENet structure to calculate the GRU extracted The weights between the feature channels are finally integrated through Linear layers to obtain the predicted field of view position.Experimental results show that the pr ediction accuracy of this method is higher than existing methods and the prediction time is short.(3)A multi-head attention field of view prediction method equipped with Convolution Block Attention Module(CBAM)is proposed.First,three consecutive convolutional layers are used to extract the features in the field of view,and then the heads are separated.Attention uses a convolutional structure,so when splitting is performed in the channel dimension,each channel retains complete information,and finally the splits after calculating the self-attention are combined,and the channel attention sum is calculated through the CBAM module Spatial attention.Experimental results show that the prediction accuracy of this method is higher than the existing representative methods.