Research On Optimization Of Quality Of User Experience Toward Adaptive Video Streaming In Real Environments

With the development of network technologies,one of the most popular Internet service,video streaming has taken a great part of network traffic with an increasing trend.Meanwhile,various video-centric services have been developed,such as video on demand,live videos,short videos,and video conferences.The demands on improving the quality of experience(Qo E)of video streaming have taken the front stage.To improve Qo E,recent studies propose adaptive bitrate(ABR)algorithms in adaptive video streaming,which select video bitrates according to network conditions to maximize the target Qo E function.The main techniques on improving Qo E are the understanding and quantization of Qo E and adaptive bitrate algorithms.Current studies provide efficient techniques for Qo E improvement in adaptive video streaming.However,they still suffer from inaccurate parameter configurations,the fixed Qo E function,and low bandwidth prediction accuracies under cellular networks.The issues make the ABR algorithms select inappropriate video bitrates in real environments,failing to adapt to Qo E functions and network conditions.From the perspective of understanding and quantization of Qo E and designing of ABR algorithms,we consider three kinds of real environments: user experience under real video platforms,diversity among users,and dynamics of cellular networks.Our work includes the following:(1)Improve overall Qo E.With collection and analysis of video data from i QIYI,we study the impacts of video quality metrics and model user engagementcentric Qo E by regression techniques.Accordingly,we define a new Qo E function and apply it in the online video system.(2)Improve specified Qo E for every user.To address Qo E diversity,we propose a parametric modeling technique and the ABR algorithm with flexible parameters which can be adjusted by users online.(3)Improve Qo E of mobile video users.We find environment-specific Markov property of network traces through the measurement and analysis of cellular networks.To improve the bandwidth prediction accuracy,we use the property for network environment identification and online bandwidth prediction.We evaluate our work by thorough experiments under real network traces and the video playback test bed,and subjective experiments of real users.The results show that our proposed(1)Qo E model characterizes the watching experiences of real users effectively,promoting 2.8 minutes of average viewing time per user(3.4%)?(2)ABR algorithm can address adaptation to Qo E diversity effectively among real users and outperform existing ABR algorithms by 3%-7% Qo E?(3)Markov-based network bandwidth prediction method improves 20%-25% prediction accuracy and 11%-20% Qo E over existing predictors.