Research And Implementation Of Real-time Comments Based HTTP Adaptive Streaming

With development of information technologies,multimedia contents have become an essential part of people’s daily entertainment nowadays.However,in scenarios of video streaming,the amount of real-time data transmission could be several orders higher than that of usual communications,meanwhile network conditions are forever fluctuating,thus providing high quality and low latency video streaming services could be a great challenge.Moreover,since perception of contents vary subjectively among viewers,how to cater services more to viewers’ tastes is another key problem requiring exploration.So as to solve problems mentioned above and deliver high quality multimedia services,plenty of researches have been addressed in this field.HTTP adaptive streaming is one of the generalized solutions.The core idea is cutting full video into segments of different bitrate levels,and choose a suitable one dynamically according to current network situations.And in this way high quality segments are delivered as many as possible,without causing unnecessary rebuffering.How to choose bitrate decision factors,when to make bitrate decisions and how to take viewers’ customized demands into consideration are difficulties when applying adaptive streaming algorithms.We propose our real-time comments based adaptive streaming and we finished the following works in the paper:(1)On the basis of viewers’ interests of video segments and taking real-time comments as raw materials,we established a more reasonable QoE model.And we comprehensively took generalized and customized factors affecting viewing experiences into consideration,and introduced normalized real-time comments weight,preferentially guaranteeing quality of video segments that attract viewers’ more interests.(2)We employed LSTM for more accurate throughput prediction,and introduced attention mechanism to avoid influences of outliers.(3)When organizing dataset,we mixed network traces collected from different environments,bringing more compatibility to trained algorithm model.We also built a system aiming to pull real-time comments from famous bullet screen video site Bilibili,easing normalization of segment weights and training and testing of the model.Comparing with state-of-the-art ABR algorithm Pensieve,experiment results show that our proposed method earns more QoE scores,without noticeably increasing player rebuffering and quality oscillation due to pouring too much network resources to highly weighted video segments.