Application-oriented Researches On Inter Prediction For Video Coding Standards

With the development of computer and communication technologies, applications based on video technologies are growing rappidly. Services such as video on demand, high definition television broadcasting, video conferencing and video surveillance are becoming more important parts of people’s daily life. In recent years, mobile com-puting and mobile communication technologies are ondergoing an explosive growth, which makes video applications be featured of the diversity of clients and transmission networks.As a crucial precondition of the widespread usage of video technologies, video coding standardization has experienced a development process of more than several decades, and produced series of video coding standards, including but not limitted to H.261, H.262/MPEG-2, H.263, H.264/AVC and HEVC, etc. Each of these standards provides a set of video coding tools, from which a compatible encoder may choose to compose qualified bitsteams. The coding tools supported by a standard largely deter-mines the coding performance that a codec can achieve.Almost all the mainstream video coding standards are based on a hybrid cod-ing framework, which contains prediction, transform, quantization and entropy cod-ing. Among them, inter prediction is the most important technique which makes use of the temporal correlation in video sequences. Two facts should be noticed about inter prediction. Firstly, the effectiveness of inter prediction has great influence on the cod-ing performance of an encoder. The effectiveness of inter prediction can be improved by promoting the accuracy of prediction signal or by reducing the overhead of mo-tion information. Secondly, inter prediction introduces dependency between different frames, which may harm the adaptation ability of a video stream. Techniques related to stream adaptation usually comes with restriction on interprediction, e.g. random access of video stream, and temporal scalability.In this paper, several application-oriented researches were performed based on ex-isting video coding standards. Firstly, an improvement for the competition based motion vector prediction scheme is proposed to reduce the motion information overhead of in-ter prediction. Secondly, the contradiction of precision of inter prediction and stream adaptation ability is studied in the context of region of interest(ROI) scalability. The contribution of this paper is three-fold.1. Based on the latest High Efficiency Video Coding(HEVC) standard, we devel-oped a novel motion vector coding scheme to achieve higher coding efficiency for motion vectors in competition based motion vector prediction framework. This coding scheme is based on decoder side reasoning and data hiding, these two tech-niques cooperate in a region based way to provide an efficient way of representing motion vector prediction index.2. Based on the rewritable H.264/SVC coding technology, a novel rewritable en-hancement layer coding method is proposed. The proposed method is featured by the ability of rewriting spatial enhancement layer and the region of interest of enhancement layer. To achieve the rewritable coding of spatial enhancement layer, transform domain residue upsampling is adopted to reduce to complexity of inter layer residue prediction of spatial enhancement layer. Besides, a rate-distortion based mode decision scheme is proposed for macroblocks in region of interest of a video sequence to take balance of coding efficienty and decoding drift for macroblocks in region of interest. Additionlly, key-picture concept is used to control the propogation of decoding drift to further improve the overall coding efficiency.3. We also proposed a novel encoding scheme for unrewritable region of interest-ing coding for H.264/SVC. Traditionally, to make region of interest be sepa-rately decodable, restriction on motion estimation process should be performed to make sure all information used for prediction is available while decoding the ROI stream. But in the context of H.264/SVC, while the background slices of enhancement layer are lost, the whole base layer information is still available, so error concealment techniques can be used to approximately produce the lost infor-mation of enhancement layer. However, using the error concealment technique blindly is not effective for ROI decoding, since the mismatch between the origi-nal reference frames and their generation approximation would introduce severe error propogation. To overcome this shortage, a double-loop encoding model is proposed. The encoder matains two sets of decoded pictures for encoding usage, one of them contains recstructed pictures of the whole enhancement layer, the other contains virtual reference pictures generated for ROI decoding separately. Motion estimation and mode decision are performed jointly considering the two encoding loop to achieve a trade-off between the two decoding scenarios.