Research On Analysis And Prediction Of User Head Movement Behavior In Panoramic Video

Nowadays,multimedia technology is evolving from providing users with a simple viewing experience to providing an immersive viewing experience.With the rapid development of the Internet,video streaming technology,as a product of the integration of multimedia technology and Internet technology,is developing unprecedentedly,and panoramic video streaming has become a popular trend in the future.Panoramic video provides users with an immersive experience.Existing panoramic video streaming still uses traditional video transmission technologies.The panoramic video can be played at the client side if the projection conversion module is added.However,panoramic video streaming faces the challenges of high bandwidth and low latency,which traditional video streaming architectures cannot cope with.Because the user's field of view is so limited that it is impossible to watch the video content of the entire spherical surface,transmitting the panoramic video in complete high-quality will cause a huge waste of bandwidth.In this regard,the ideal solution is to predict the user's future head position,and only transmits the video content at high quality for the surface that a user's future viewport covers.However,prediction errors are inevitable.In order to prevent the Qo E degradation caused by prediction errors,for example,the blank blind area caused by the sudden head movement,it is reasonable to transmit high-quality video within the predicted viewport,and low-quality video outside the predicted viewport.In this way,it is possible to achieve good balance between saving bandwidth and maximizing quality of experience.Therefore,a more accurate head movement prediction can effectively reduce bandwidth and improve user's Qo E.With the development of artificial intelligence,people can use a large amount of data to train head movement prediction models and obtain more accurate prediction results.However,head movement prediction also has certain challenges.The first one is of data sets.Currently,the data sets are not large enough to allow useful model training.In addition,differences in user preference,video content,viewing experience,environment,viewing purpose and so on will cause differences in user's viewing behavior,which makes head movement hard to predict.The contributions of this thesis are as follows:(1)A head movement data set and a saliency data set were established;(2)Head movement analysis on the data set were conducted,and the differences in user behaviors under restricted/unrestricted movement conditions were investigated.What's more,user behavior was clustered,and the transferring characteristics of the user's head movement were analyzed;(3)Combining historical head movement data with cross-user data,a head movement prediction algorithm based on Markov chain was proposed,which referred to the head movement information of the user and the head transferring information of other users to predict the user's head movement.In addition,to cope with the inaccurate prediction caused by insufficient cross-user data information in the algorithm,an improved Markov-chain based head movement prediction algorithm was also proposed,which effectively improved the accuracy of head movement prediction.