Statistical Characteristics and Model of MPEG VBR Video Stream

MPEG VBR video stream cannot be well modeled because of its complexity. Many papers on VBR video modeling are concentrated on Markov-modulated scene models. One of them is based on GOP (Group of Picture), but for the sake of so large state space of Markov model, it greatly increases computational complexity, so it is not easy to be implemented in real-time environment; Another approach is the scene-clustering algorithm based on Frame Bit Rate, yet because of the periodic correlation of frames, it is still hard to fit the distribution and autocorrelation functions of video stream sequence. We analyze the statistical characteristics of MPEG VBR video stream in this paper. Our research shows, by clustering the video sequence into independent classes, the state space of Markov chain can be greatly decreased, so all classes can be modeled by a Markov modulated chain. In addition, individual class can be modeled by TES model based on GOP bit rate in order to avoid the periodic correlation of frames. The first order (distribution) and second order (autocorrelation) statistical characteristics of scenes in every individual class can be easily fitted by Gamma function and double exponential function, which even more simplifies the modeling complexity.Our research result is useful to multimedia flow modeling, network performance analysis, B-ISDN design and network control strategy such as congestion control, call admission control (CAC), dynamic multiplexing, etc.