Distributed Video Coding for Multiview and Video-plus-depth Coding

Matteo Salmistraro

Research output: Book/ReportPh.D. thesisResearch

488 Downloads (Pure)

Abstract

The interest in Distributed Video Coding (DVC) systems has grown considerably in the academic world in recent years. With DVC the correlation between frames is exploited at the decoder (joint decoding). The encoder codes the frame independently, performing relatively simple operations. Therefore, with DVC the complexity is shifted from encoder to decoder, making the coding architecture a viable solution for encoders with limited resources. DVC may empower new applications which can benefit from this reversed coding architecture. Multiview Distributed Video Coding (M-DVC) is the application of the DVC principles to camera networks. Thanks to its reversed coding paradigm M-DVC enables the exploitation of inter-camera redundancy without inter-camera communication, because the frames are encoded independently.
One of the key elements in DVC is the Side Information (SI) which is an estimation of the to-be-decoded frame. Another key element is the Residual estimation, indicating the reliability of the SI, which is used to calculate the parameters of the correlation noise model between SI and original frame. In this thesis new methods for Inter-camera SI generation are analyzed in the Stereo and Multiview scenarios. Furthermore on-line correlation noise models are proposed. On-line models are needed to enable the codec to be used in realistic scenarios. Focus is put on developing and investigating robust fusion techniques, able to correctly fuse various Sis. Learning algorithms for improving the fusion procedure are also explored in this work.
Optical Flow (OF) is a powerful motion estimation technique, enabling precise and flexible calculation of a Motion Vector (MV) for each pixel of the frame. The high density of MVs has discouraged the use of OF in conventional predictive coding, because of the need to code the MVs. On the other hand DVC can exploit OF because the Motion Estimation (ME) is only performed at the decoder. In this thesis it is proposed to use OF for joint disparity and motion calculation in M-DVC and for joint motion estimation in texture and depth frames in video-plus-depth.
Rate Adaptive (RA) error correcting codes are the core of all the modern DVC codecs, nevertheless they suffer from efficiency problems, most notably for short block lengths and high correlations between SI and original signal. A novel coding architecture based on RA BCH (Bose-Chaudhuri-Hocquenghem) codes has been presented, along with an analytic model for predicting its performance and many different methods to improve the reliability of the decoded results.
Original languageEnglish
PublisherTechnical University of Denmark
Number of pages194
Publication statusPublished - 2014

Fingerprint

Dive into the research topics of 'Distributed Video Coding for Multiview and Video-plus-depth Coding'. Together they form a unique fingerprint.

Cite this