A monograph on prediction models for MPEG encoded VBR video traffic

Abstract

This work focuses on real-time predictions of MPEG encoded VBR traces. These experimental traces were obtained from publicly available online video trace library. the predictor considers only the local traffic characteristics and ignores the average traffic characteristics. This thesis work addresses this issue by exploiting the importance and simplicity of MA predictors. This work addresses the issue of capturing the burstiness property by proposing more accurate predictors for I-frame traffic (as they are the direct representatives of burstiness as well as scene-changes) particularly during scene-change transitions. This work proposes to improve prediction accuracy while reducing predictor complexity during scene-change transitions for real-time traffic. Scene change is characterized by a fuzzy membership function and based on the function value, the step-size variable is adjusted. This is followed by a little training iterated over a few steps for refining the weight vector of the NLMS algorithm. Normalized LMS predictor is used as baseline predictor in this work. A unique contribution of this work is to address this issue of improving the prediction performance of training based models by refining the training dataset. A FIS is used to determine the degree of noise in a training instance. The current work has successfully addressed this issue of reducing fluctuation in the prediction error by deploying a simple algorithm in conjunction with adaptive real-time NLMS predictor. This work proposes a new multiplexing methodology, for VBR video sources, for improved SMG. The proposed method called ERA not only ensures near zero I-frame collisions, but also equals in complexity with existing frame-aligned technique, while showing improved performance. The work reported in this thesis makes use of single layer framesize H.264 traces for most of the experiments, and performance analysis is done using measures like SNR-1, RPE, and RMSE. newline newline newline