| As major video websites attach importance to the field of virtual reality and users’ demand for viewing panoramic videos online,panoramic video as an emerging form of multimedia has received widespread attention and will usher in explosive growth.However,compared with traditional video,panoramic video needs to transmit a large amount of data,and the existing mobile network cannot meet its demand for bandwidth.On the premise of ensuring the viewing quality of users,how to design a transmission mechanism to optimize the transmission of panoramic video has become the research focus of breaking the bottleneck of panoramic video business development.Faced to panoramic video live broadcast and on-demand scenarios,this thesis studies the real-time transmission optimization of panoramic video.The main research contents are as follows:(1)The transmission scheme based on the viewing angle can effectively reduce the amount of panoramic video transmission data and reduce the bandwidth occupation,but at the same time,it puts forward requirements for accurately predicting the user’s viewing angle.This thesis proposes a method of using video pixel motion feature information and users’ historical viewing data to predict the viewing angle,which improves the accuracy of viewing angle prediction.(2)For the live broadcasting of panoramic video,this thesis proposes a pixel-level FOV(Field Of View)extraction scheme that combines human visual characteristics and panoramic video projection format,which effectively reduces the bandwidth occupation while transmitting highquality FOV images.For multi-user scenarios,this thesis proposes a perspective clustering algorithm by grouping users with similar perspectives into the same multicast group to obtain multicast gains.Experimental results show that this scheme can effectively reduce the data during panoramic video transmission and save network bandwidth.(3)For the on-demand service of panoramic video,the transmission scheme based on the spatial block(Tile)is the current mainstream optimization scheme.This thesis uses the reinforcement learning method to study the selection of tile rate,which can reduce the impact of network status on user immersive experience and improve user experience quality.First,construct the user experience quality metric as the reward function of the reinforcement learning network,and secondly use the A3C(Asynchronous Advantage Actor-Critic)algorithm to train the adaptive model based on this reward function to maximize the long-term user experience quality.This thesis further uses multi-agent reinforcement learning algorithm to improve the fairness of experience quality for multiple users.The comparative experiment results verify that the proposed scheme can provide all users with a better quality of experience. |
