A genetic approach to Markovian characterisation of H.264 scalable video

Abstract

We propose an algorithm for multivariate Markovian characterisation of H.264/SVC scalable video traces at the sub-GoP (Group of Pictures) level. A genetic algorithm yields Markov models with limited state space that accurately capture temporal and inter-layer correlation. Key to our approach is the covariance-based fitness function. In comparison with the classical Expectation Maximisation algorithm, ours is capable of matching the second order statistics more accurately at the cost of less accuracy in matching the histograms of the trace. Moreover, a simulation study shows that our approach outperforms Expectation Maximisation in predicting performance of video streaming in various networking scenarios.