RIA: Object-Oriented Motion Decomposition and Estimation with Application to Low-Bit-Rate Video Coding

本课题研究时变图像序列中的运动估计及其在视频压缩中的应用。所提出的方法基于变换域运动表示，该变换域运动表示通过一组变换系数来描述运动对象的空间位移场，每个变换系数指定基本运动对表观运动的贡献。对于包含具有混合运动的多个对象的图像，使用这种面向对象的运动表示的困难通过多分辨率四叉树结构中的联合运动估计和对象分割过程来克服。拟议的项目由两部分组成。第一部分包括以下几个方面的理论和算法研究：1)可行的运动变换基础；2)变换域运动估计；3)分层联合运动估计与目标分割；4)运动补偿时间内插。第二部分利用第一部分中提出的技术对低比特率面向对象视频编码进行了研究。第一部分的结果将有助于图像序列处理和分析的理论，并将为视频信号处理、三维场景分析和目标跟踪中的几个关键问题提供新的解决方案。特别是第二部分的成功，将为低比特率视频编码提供新的替代方案，低比特率视频编码对于窄带信道上的视频编码和视频数据库应用是必不可少的。

This project considers the estimation of motion in a time-varying image sequence, and its applications in video compression. The proposed approaches are based on a transform domain motion representation which describes the spatial displacement field of a moving object by a set of transform coefficients, each specifying the contribution of a basic motion to the apparent motion. The difficulty with using such an object-oriented motion representation for images containing multiple objects with mixed motions is countered by a joint motion estimation and object segmentation procedure in a multi-resolution quad-tree structure. The proposed project consists of two parts. The first part includes theoretic and algorithmic investigation on, 1) viable motion transform bases, 2) transform domain motion estimation, 3) hierarchical, joint motion estimation and object segmentation, and 4) motion compensated temporal interpolation. The second part encompasses the studies on low bit rate object oriented video coding using the techniques produced in the first part. Results from the first part will contribute to the theory of image sequence processing and analysis, and will provide new solutions to several key elements in video signal processing and three dimensional scene analysis and object tracking. The success of the second part, in particular, will suggest new alternatives to low bit rate video coding which is indispensable for video coding over narrow band channels and video database applications.