Researches On Fast Algorithm For Multi-view Depth Video Coding

With the development of three-dimensional (3D) display and related technologies,3D videoapplications, such as3DTV and free viewpoint TV (FTV), have been booming up recently.Multiview plus depth (MVD) data format meets the requirements of3D video system and supportswide multi-angle for3D displays. Hence, MVD is the main3D representation format. MVDsignals consist of multiview color video and multiple associated depth video. Depth video,representing the relative distance of each video object to the recording cameras, mainly aids inrendering virtual views by depth-image-based-rendering (DIBR) technique. In MVD systems, bothcolor videos and depth videos need to be compressed prior to transmission. The data amount ofMVD is tremendous because it is proportional to the number of cameras. The large amount ofMVD data requires huge storage space and wide transmission bandwidth. Therefore, it is necessaryto develop an encoding method with high compression efficiency and low complexity for real-timevideo applications. In3D video (3DV) system, depth video and color video share a codec and adecoder. Joint Multiview Video Coding (JMVC) uses some techinologies, such as searching allmodes, multi-reference frame selection technique, distortion (RD) optimization technique, motionestimation, disparity estimation and so on, to obtain high compression efficiency. However, thesetechniques lead to increasing computational complexity. This paper mainly commits to fastalgorithms for multiview depth video coding. And the contributions of this paper are:(1) This paper proposed a depth segmentation method. People are usually interested in regionswith large depth value and depth discontinuous regions. Moreover, the edge of objects is moreimportant to DIBR than other regions. Considering depth perception and the influence of depth toDIBR, the depth video is segmented into three regions: edge, flat foreground and flat background.Based on the results of depth segmentation, two fast algorithms for depth video coding areproposed in this paper. Firstly, this paper proposes a fast mode decision algorithm for depth videocoding based on depth segmentation. Mode distribution of different regions and mode correlationbetween neighboring MBs are analyzed in detail. And then encoding time, bit rate and videoquality of virtual view of the proposed algorithm are tested. Experimental results show that theproposed algorithm saves encoding time ranging from82.49%to93.21%with negligible qualitydegradation of rendered virtual view images and bit rates increment. Compared with otherexcellent fast algorithms for depth video coding, the proposed algorithm is more superior. Secondly,this paper proposes a depth segmentation based fast multi-reference frame selection algorithm fordepth video coding. Based on the characteristics of multi-reference frame selection for differentregions, different regions adopt different optimization strategies of multi-reference frame selection.The experimental results show that the proposed algorithm reduces the encoding time ranging from 34.54%~56.91%while it keeps video quality of virtual view and rate distortion performance ofdepth video coding.(2) This paper proposes a fast mode selection algorithm for multi-view depth video codingbased on correlation. Firstly, a mode complexity factor is defined. According to the factor, depthvideo is divided into two regions: simple model region and complex model region. Then,correlation characteristic and macroblock mode distribution for the two regions are discussed andanalyzed in detail. Finally, the fast algorithm is presented based on the analysis. The experimentalresults show that the proposed algorithm saves encoding time ranging from65.57%to92.72%while it maintaining high virtual rendering and compression performance.