Multiview video coding for 3D telepresence

In this thesis, we develope several novel algorithms and two complete encoders for multiview video coding (MVC). The major contributions can be summarized in three aspects. In particular, we first propose an edge-preserving regularization scheme to calculate either 1D disparity fields or 2D motion fields. After confirming its performance by comparing with the existing algorithms, the separate regularization scheme is extended to a joint estimation scheme that calculates two disparity fields and two motion fields for two successive image pairs simultaneously. Secondly, we develop an MPEG-4 compatible multiview video encoder which integrates with the joint disparity and motion estimation scheme. Besides, various aspects of the encoder are investigated, including a comparative study of several view-level frame prediction structures and the rate control algorithm for multiview video coding. Thirdly, we propose a framework of scalable MVC using 4D wavelet. The wavelet-based multiview video coding system is more flexible and extendable, and it can provide temporal, spatial, SNR (Signal-to-Noise-Ratio) and view scalabilities. In the wavelet MVC scheme, general lifting structure is introduced, based on which we propose a disparity compensated view filtering (DCVF) for wavelet decomposition along the view direction. We also propose a flexible decomposition structure based on the analysis of the temporal correlation and the view correlation. In addition, we develop the entire codec, including the design of the macroblock coding modes, subband coding, rate control, etc. The proposed scheme overcomes a variety of limitations exhibited by existing methods and provides a natural and elegant solution to MVC.DOCTOR OF PHILOSOPHY (SCE