Statistical modeling for networked video: Coding optimization, error concealment and traffic analysis.

摘要

Video coding has attracted much attention in the recent past, especially due to the large amount of digital video content available today. Video transmission and storage requirements result in efficient compression techniques with many different evolving compression standards, such as H.263 and MPEG-4. Besides efficient compression, video coding techniques have to ensure good video quality while involving real-time processing. Complexity, quality and bit rate are factors that measure success of a video coding scheme. The focus of this thesis is on optimizing the video coding process to improve performance in terms of one or more of these three factors. We use statistical modeling techniques to achieve this optimization goal. Specifically we examine two parts of the video coding process, mode decisions and error concealment. Mode decisions involve selecting the optimal modes of operation under certain constraints. We build a classification based framework for making mode decisions to minimize the coding cost that may be defined in terms of the three parameters, complexity, quality, and bit rate. We propose a scheme for model based error concealment, i.e., using a statistical model for the region of interest to replenish any data lost due to errors in network transmission. We introduce a new and efficient statistical model called Mixture of Principal Components (MPC) to capture the properties of the region of interest. We show that this model is more efficient than the traditional Principal Component Analysis (PCA) in capturing data variations, especially when the data consists of samples distributed in multiple clusters. We also use this model for an example face recognition task in order to highlight some other applications for this general statistical framework. We realize that both the mode decisions as well as the error concealment optimizations require feedback from the network regarding the available bandwidth, loss probability, and delay. Hence, in the last part of this thesis we focus on modeling the variable bit rate video traffic so that we may use this traffic to probe the network to determine the network condition and optimize our coding algorithms appropriately.