Abstract
As a rate and format conversion process, video transcoding is getting more and more important in many video communication applications. In this thesis, we investigate a series of practical video transcoding techniques for multipoint video conferencing. Firstly, several single-point and multipoint video transcoding architectures are investigated and compared in terms of PSNR performances and implementation complexities.We present a top-down dynamic rate control method in multipoint video transcoding to enhance the visual quality of the active frames, the sub-windows of active conferees and the face regions which are often the areas of interest. An activity-based frame-layer bit allocation scheme is proposed to reduce the picture quality fluctuation due to the variations of picture complexity thus the flickering effect will also be reduced. In a video conference, there are in general only one or two active conferees at one time which attract the attention of the conferees. The proposed sub-window rate control method firstly identifies the active conferees from the multiple incoming video streams by calculating the temporal and the spatial activities of the conferee sub-windows. In our method, the sub-windows of inactive conferees may be dropped and the saved bits are reallocated to the active sub-windows by using a rate-distortion optimized bit allocation approach based on the H.263 TMN-8 rate-distortion models. This bit rate reallocation can significantly enhance the visual quality of the active sub-windows, while the temporal resolution of the inactive ones is sacrificed, which is, however, relatively invisible.In addition, since face region is usually the center of focus in video telephony applications, we also propose a novel skin-color based scheme to locate and track the face regions in real-time, and then classify the macroblocks into face and non-face classes accordingly. A face-assisted dynamic distortion weighting adjustment scheme is subsequently used for macroblock-layer bit allocation to enhance the visual quality of the face regions.We also present several motion vector composition schemes which can be used to compose the unavailable motion vectors in the dropped frames due to limited bit-rates or frame-rates of the user clients in video transcoding. Of the presented methods, we suggest to use the forward dominant vector selection scheme (FDVS) scheme because it can provide reasonably accurate approximation of motion vectors with lowest computational cost and memory requirement.The experimental results show that the proposed methods can significantly improve the PSNR performance and the subjective quality on the active sub-windows and the face regions with low extra computational complexity, while the degradation introduced on the inactive sub-windows and the non-face regions is relatively insensitive to human perception. Moreover, the proposed algorithm is fully compatible with the H.263 video coding standard.We finally show the implementation of an H.323 standard compliant video conference system prototype which adopts the proposed video transcoding techniques in the MCU server. The proposed H.323 MCU server can also serve as a gateway between H.323 LAN user terminals and H.324 WAN user terminals. We also present an object-based transcoding scheme using chroma-key techniques and show how to provide user friendly personal presence controllable features in H.263 compliant video conference systems accordingly.