Dissertation
Dissertation > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device

Research on Depth Estimatipn and Video Coding Technologies in Free-Viewpoint Video

Author LiuXiaoXian
Tutor ChangYiLin
School Xi'an University of Electronic Science and Technology
Course Communication and Information System
Keywords Free-viewpoint TV Free-viewpoint Video Depth Quantization Depth Estimation Virtual View Synthesis Motion Vector Prediction
CLC TP391.41
Type PhD thesis
Year 2011
Downloads 0
Quotes 1
Download Dissertation

Free-Viewpoint Video (FVV) system is an innovative visual media that allowsviewing a3D world by freely changing the viewpoint. FVV will bring an epochalchange in the history of visual media since such a function has not yet been achieved byconventional media technology. In the sender of FVV system, it is required to capturemultiview video, which results in huge amount of original data in FVV. Thereby,efficient FVV compression technique is crucial for its wide applications. To provide theFVV within a certain range in the receiver of FVV system, it is necessary to synthesizevirtual view using the depth-image-based-rendering (DIBR) technology based onmultiview video and corresponding depth maps. Thus, generating depth data with highquality is one of the key enabling factors for FVV system realization.This dissertation investigates key techniques in FVV. Major contributions of thisdissertation are summarized as follows:1.An adaptive non-uniform quantization method is proposed for depth data inFVV system, with which the distortion due to depth data quantization can be reduced.Based on the statistical distribution characteristics of the depth data, the proposedmethod firstly determines the valid depth sub-spaces. Then, such depth sub-spaces arefinely divided to derive the adaptive quantization look-table. Finally, the original depthdata are adaptively quantized using this quantization look-table. Experimental resultsdemonstrate that the adaptive non-uniform quantization method can significantly reducethe quantization distortion of the depth data, which benefits the high-accuratereconstruction of the depth data and the better subject visual effect of the synthesizedimage at the receiver of the FVV system.2.A novel depth characteristic of edge pixels based depth estimation method(DCE-SADE) is proposed to improve the depth estimation precision at the object edges.In the proposed method, based on the depth characteristics of the object edges, thepixels in the image are firstly divided into three categories, and then the correspondingsmoothing terms of the matching cost function are derived. The experimental resultshave shown that compared with the existing edge information based semi-automaticdepth estimation method (E-SADE) in the FVV system, the proposed method cangreatly improve the depth estimation precision at the object edges, and the PSNR of thesynthesized virtual view based on the proposed method can be improved to a differentextent with an average of0.32dB at the receiver of the FVV system. 3. A texture video-assisted motion vector predictor for depth maps coding isproposed to improve coding efficiency for depth map. Based on the analyses of motionsimilarity between texture videos and their corresponding depth maps, the proposedapproach uses the motion vectors of texture videos and the median predictor jointly todetermine the optimal predicted motion vector for depth maps coding by employing arate-distortion (R-D) criterion. Experimental results demonstrate that compared with themedian predictor utilized in H.264/AVC, the proposed method can save the maximumand average bit rate as high as4.89%and3.68%, respectively, while guaranteeing thequality of synthesized virtual views.

Related Dissertations
More Dissertations