A new prediction model for MPEG coded video: two-sided Markov-renewal model (TSMR)

This paper introduces a new prediction-modeling framework for VBR video. The idea is to model the durations of the variations in the source traffic using a modified Markov-renewal process. These variations correspond to taking the first difference in each of the decomposed (I, P, B, and GOP) frame size process, which has the effect of transforming it into a stationary one. The resulting Markov states can be classified into two groups: low-variation and high-variation. A low-variation state corresponds to a small difference between adjacent frames or group of frames, whereas a high-variation state corresponds to a significant change in size. The values of the variations are then modeled to match both the autocorrelation structure and marginal distribution function. The resulting model is called the two-sided Markov renewal model (TSMR) and is designed specifically for prediction. A simple Markov decision policy is developed and used for this purpose. In order to evaluate the model, real-time predictions are carried out on a number of MPEG coded video streams, simulation results are discussed and compared with their empirical counterparts. The model is parsimonious in terms of parameters used and memory required. Computation of the prediction algorithm is fast and simple. Only minimal knowledge of the source traffic is required to drive the predictor machine. It is most suitable for the task of dynamic bandwidth allocation In which only very little knowledge about the source is available in advance.