Research On The Method Of Streaming Media Live Broadcast Transmission Rate Adaptation

The application of network streaming media is one of the hotspots in the era of mobile Internet.The client uses the network streaming media player to download the video files in the form of "blocks" in order.Generally,in order to improve the quality of user experience and make the user have a smooth and high-definition video viewing experience,the existing players use the adaptive bitrate(ABR)algorithm to adjust the bitrate of the video played by the user.The scheme of bitrate adjustment is the core problem to be solved in current streaming media technology.Although a large number of bitrate adjustment schemes have been proposed(for example,BB and Bola based on buffer,RB based on throughput prediction,and MPC combining buffer and throughput prediction),all of them have a common disadvantage: Based on fixed rules or using simple models.Because the network environment is very complex,network traffic and bandwidth will change at any time,and the existing schemes can not completely deal with the changing network environment,so each method can not achieve the best decision in all environments.To solve these problems,based on the analysis of the advantages and disadvantages of the most popular ABR algorithms based on throughput prediction and buffer,this paper proposes a new ABR algorithm model on the basis of Pensieve,an ABR algorithm model based on deep reinforcement learning.In order to solve the problem that the length of the chunks in the video data set is not the same,the model processes the video data set to make the block length reach the preset value,so that the model can support more video data sets.Then,the model uses convolution neural network to extract the features of the processed data set,and then the extracted features are fused.Finally,the fusion features are processed by LSTM,which greatly improves the performance of the model.Experiments show that the QoE of the propoesd ABR algorithm model based on deep reinforcement learning is higher than other models,about 17% higher than the second highest model RB,and about 19% higher than the similar model Pensieve.