Advanced Bitstream Switching Techniques for H.264 Video Streaming and Multi-View Video.

摘要

The goal of video streaming is to transmit pre-encoded videos from a server to a client over the Internet. Owing to the large bandwidth fluctuation, efficient (11 adaptation to channel bandwidth is critical when transmitting video over various networks. The most straightforward way is to represent each pre-encoded video sequence by using multiple and independent bitstreams of different bitrates and quality. An efficient switching mechanism among different bitstreams is then used to select an appropriate bitstream. Therefore, some novel techniques are suggested for bitstream switching with the minimum storage requirement and the reasonable complexity of the server.;In H.264, the new frame type - Synchronization-Predictive frames (SP-frames) - is developed for multiple bitrate streaming with the support of seamless switching. However, the trade-offs are the bulkiness of SP-frames and degradation of quality. Our analysis reveals that the conventional pixel-domain motion estimation is not appropriate for encoding secondary SP-frames because a significant amount of additional storage is induced. For this reason, we propose a new motion estimation and compensation technique, which is operated in the quantized transform (QDCT) domain instead of pixel-domain, for coding secondary SP-frames. The encoding structure of secondary SP-frames is re-designed. Our proposed work keeps the secondary SP-frames as small as possible without affecting the size of primary SP-frames. Simulation results demonstrate that the size of secondary SP-frames can be reduced remarkably.;In the meantime, multi-view video service has been attracting more and more attention recently. Different from the traditional single-view video, multi-view video brings in a brand new viewing experience with a high degree of user interactivity. It allows users to select their favourite viewpoint. It switches the bitstream at a particular view when necessary. The SP-frames, with capability of drift-free switching, can be directly employed in the low-delay viewpoint switching of multi-view videos. It means that QDCT-domain motion estimation can then be proposed to apply on the multi-view video scenarios. Experiments show that the weak interframe correlation in multi-view videos, comparing to multiple bitrate videos, can get benefit from our proposed scheme due to the fact that QDCT-domain motion estimation aims at minimizing residues rather than finding true motion vectors.;Although the new QDCT-domain scheme exhibits promising results for reducing the storage requirement of secondary SP-frames in multiple bitrate video and multi-view video systems, there are still some cases that pixel-domain motion estimation outperforms QDCT-domain motion estimation. In our work, we propose two hybrid motion estimation algorithms (frame-based and macroblock-based) to select an appropriate domain of motion estimation in secondary SP-frame coding. The selection is based on inter-frame correlation, which is measured using the bit-counts of its corresponding primary SP-frame. Results show that the proposed algorithm significantly decreases the size of secondary SP-frames and consumes much less time.;By employing QDCT-domain motion estimation, the work in this thesis shows outstanding improvements in terms of size and computational complexity for both multiple bitrate video and multi-view video applications. There is no doubt that the results of our work will certainly be useful for the future development of bitstream switching systems.