Statistical Characterization VBR MPEG at the Slice Layer

We studied the ATt\1 cell-generation statistics of three VBR encoded i\1PEG video sequences at the Slice layer. Each of the video sequences contained Intracoded (I)~ Predictive-coded (P) and Bidirectional-coded (B) frames and included special camera effects such as zooming, panning and sudden scene changes. Each sequence consists of 150 frames at a rate of 30 fps. Previous research characterized sequences which did not include B frames. This was due to its long encode time making it unattractive for real-time video applications. We examined video sequences with B frames since they offer the highest amount of COinpression and increased bit-rate variation which can be exploited when multiplexing multiple video streams, Also, it is likely that applications where video is stored for retrieval will contain B frames. The motivation behind studying the cell generation process at the Slice layer is that the slice size makes it well suited for transport packet payloads. It would be feasible to have ATMConvergence Sublayer (CS) payloads consisting of slices where destination nodes can detect slice errors [U1d provide corrective action or attempt error concealment. Also, characterization at the Slice layer captures the spatia] behavior of each frame, We calculated quantitative metrics such as compression ratio and picture density to compare the video sequences and discuss the effects of content on these metrics, We determined that the compression ratio can differ by a factor of 5 due to content. We showed the macroblock type distribution for each sequence based on frame type and discussed their effects on compression ratio. We gave the distribution of cells/slice for each sequence and determine if the gamma distribution is a good fit by using QQ plots. We found that none one of the sequences fits the gamma distribution well. This makes it unlikely that a single distribution can apply to all IIB/P video streams, We noted that at the Slice layer. the data is highly correlated with negative linear decay. Also, peak-to-mean ratios varied from 3.~ to 6.2 and the coefficients-of-variation varied from.v? to 1.09. We pointed out that dynamic bandwidth allocation schemes which request bandwidth in fixed symmetric increments might not be appropriate for sources with I/B/P frames. We presented a multiplexing scheme whereby video streams are time shifted in order to overlap B frames from one stream with I and P frames of another. We found that the peak-to-mean ratio and coefficient-of-variation is reduced and that the multiplexed output stream fits the Gamma distribution well, Finally, we used the Markov Chain to model the cell generation process and show that an 8 state Markov Chain is able to match the cells/slice distribution well.