Scene change detection-based Discrete Autoregressive modeling for MPEG-4 video traffic

With more than 16 billion videos streamed on YouTube during last May and recent estimates by Cisco that mobile video traffic will increase 25-fold between 2011 and 2016, there is a pressing need to adequately serve large numbers of simultaneous online video transmissions. Network providers need to be able to guarantee the strict Quality-of-Service (QoS) requirements of real-time variable bit rate (VBR) video users, and a good statistical model for multiplexed video traffic can help significantly to evaluate and enhance network performance under various video loads. In this paper, we propose and evaluate a new hybrid video traffic model for MPEG-4 video. The genres of the videos considered in our study include lectures, cartoons, talk shows, action movies, sci-fi and sports. In the first part of our work, we build a Discrete Autoregressive model of order one (DAR(1)) and discuss its efficiency on capturing the behavior of real MPEG-4 video traces. We then proceed to build and evaluate a hybrid model which combines the DAR(1) model with a scene-change detection and classification algorithm, in order to provide us with higher modeling accuracy. The classification is performed based on the average number of bits generated during the scenes and the scene activity is modeled by a Markov chain where each state represents the degree of activity (high/low). Our extensive experimental evaluation study shows that the proposed hybrid method significantly improves the efficiency of pure DAR(1) schemes.