Research On Multiuser Panoramic Video Transmission Model Based On Machine Learning

As a combination of traditional video and virtual reality technology,panoramic video is becoming increasingly attractive to researchers and industry.A panoramic video captures the whole scene in the space,resulting in a huge amount of data.This makes the streaming and delivery of panoramic video data much more difficult than traditional video.Recently,researchers have proposed solutions to one-on-one panoramic video delivery which are based on viewport adaptive streaming.These solutions consider the fact that audience can only see the part of video frame which is inside his/her viewport.Instead of fetching the entire video data,these solutions will only transmit the visible portion of the video to save bandwidth.However,as for the case of multiuser panoramic video streaming,existing solutions are not aware of the similarity between audience's viewport,which leads to redundant transmission.In this paper,we propose a multiuser panoramic video transmission model based on machine learning.This model will exploit the similarity between audience's viewport and optimize the bandwidth consumption by multicasting the conjoint viewport.More specifically,it uses machine learning models to predict the head movement of audience as well as their viewport overlaps.Based on the prediction,it will divide the visible portion of the video frame into parts which are visible to only one user and parts which are shared by multiple user.The unique parts will be transmitted using unicast as usual,and the shared parts will be transmitted using multicast to avoid redundant transmission and save bandwidth.To prove the efficiency of our model,we conducted tests and simulations using real head movement data of panoramic video audience.Results show that the proposed model can accurately predict the head movement and conjoint viewport of users,and can reduce the global bandwidth consumption by 10% to 40%.