【作者】 史立；

【导师】 张兆扬；

【作者基本信息】 上海大学 ， 通信与信息系统， 2001， 博士

【摘要】 MPEG-1、MPEG-2、H.261和H.263等经典的视频编码标准由于不需要对场景进行分割或分析，所以从某种意义上说是低级处理技术。尽管它们可以获得比较高的压缩比且应用非常广泛，然而随着多媒体和基于内容交互性的应用日益增多，需有新的视频编码方案。 正在兴起的MPEG-4视频编码标准由于引入了视频对象平面(VOP)的概念而具备基于内容的交互功能。由于输入视频序列的每一帧被分割成任意形状的视频对象平面(VOP)，这样每个VOP描述了一个语义意义的对象或所感兴趣的视频内容。 将视频序列分解成VOP是一项相当困难的任务。生成VOP的难点是所感兴趣的对象往往难以用诸如彩色、亮度和光流等传统的图象分割特征来表达，所以经典的分割算法无法获得有语义的划分。 本论文探讨了生成视频对象平面(VOP)的技术，并且提出了从视频序列中自动分割VOP和自动提取VO的算法，由于视频对象通常以异于背景的运动来表征，所以本算法结合运动信息进行处理。 理论上可以使用光流或运动场估计进行分割，但其对噪声相当敏感，且分割精度也会因遮挡和孔径效应的影响而剧烈降低。另一方面，可通过变化检测掩模将遮挡区域标识为变化的，但除非对象包含足够的纹理信息，否则就难于精确定位对象位置，而且还必需由附加装置来填充对象内部的洞隙。20()1年上海大学博士学位论文 本文提出一套分割算法可有效地分割出视频对象平面(VOP).主要有三种方法:一是基于形态运动滤波和活动轮廓模型的分割方法，另一是基于运动对象边缘帧差和豪斯道夫对象跟踪的分割方法.此外，为了解决视频对象仅有局部的偶尔运动，用形态运动滤波器和边缘帧差图象无法提取出整个初始语义对象的间题，本文还提出一种基于帧差边缘和灰度边缘之间高度相关性，通过累积运动信息得到语义对象二值边缘模型而分割前景对象的Vop分割算法. 其中第一种方法，首先根据形态运动滤波器提取出初始对象，滤波准则为全局支配运动与局部运动的差.标识出初始对象后，再阑值化和去噪，就可以提取出初始运动对象和其二值轮廓模型.然后再使用活动轮廓模型在后继帧中跟踪和匹配位置和形状都发生了变化的对象，最后根据一系列二值轮廓模型，就可以将序列中的视频对象提取出来.此算法特别适用于前景对象运动不甚剧烈，静止或运动背景的序列. 第二种方法首先根据边缘帧差图象标识出前景对象，再使用Calmy算子得到此前景对象的二值边缘模型，为了容纳对象在后继帧中的位置变化，使用经本文改进的豪斯道夫距离跟踪器跟踪此对象，以修改的二值边缘模型适应对象在后继帧中的形变.最后由一系列二值边缘模型引导提取出视频对象.此算法特别适用于前景对象的运动较激烈和非刚性形变较大而背景是静止的序列. 使用形态运动滤波器或变化检测掩模无法标识轻微运动的语义对象，例如某些视频会议中的头肩序列的对象.因此我们提出一种基于运动对象边缘和帧差边缘高度相关性的双向运动估计跟踪算法，且修改运动区域边缘的空间位置，删除先前帧中帧 视频对象平面分割和视绷对象提取方法的研究差图象的边缘，增加当前帧的帧差图象的边缘，从而形成语义对象的边缘图象，再根据一系列边缘图象提取出运动对象. 本文所提出的算法基于模式识别和对象跟踪原理，因而避免了某些与运动估计相关的间题.实验也证明本文所提出的算法可以处理许多典型的视频序列.最后本文也指出对多视频对象提取和遮挡情况下视频对象提取的问题尚需作更深人探索，才能有效解决. 本文课题是上海市自然科学基金(卯2D14035)的资助项目.更多还原

【Abstract】 Traditional video standards such as MPEG-1, MPEG-2, H.261 or H.263 are low-level techniques in the sense that no segmentation or analysis of the scene is required. They can achieve high compression ratios and are suitable for a wide range of applications. However, with the increasing popularity of multimedia applications and content-based interactivity, new video coding schemes are necessary.The standard MPEG-4, which is currently being developed, enables content-based functionalities by introducing the concept of video object planes (VOP’s). Each frame of the input sequence is segmented into arbitrarily shaped image regions (VOP’s) such that each VOP describes one semantically meaningful object or video content of interest.Decomposing a video sequence into VOP’s is very difficult in this field. An intrinsic problem of VOP generation is that objects of interest are not homogeneous with respect to low-level features such as color, intensity, or optical flow. Thus, conventional segmentation algorithms will fail to obtain meaningful partitions.This paper addresses video object plane generation and presents a new algorithm that can automatically extract moving objects from a sequence. Since these objects are characterized by a different motion from that of the background, some type of motion information must be incorporated into the segmentation algorithm.Optical flow or motion fields could theoretically be used, but they are extremely noise sensitive and their accuracy is limited due to the aperture and occlusion problem. Change detectors or difference images on the other hand mark occlusion areas as changed,while the objects themselves are unchanged unless they contain sufficient texture. This makes exact boundary location difficult and an additional mechanism is necessary to fill the holes inside objects.In this thesis we propose a set of algorithms which can efficiently segment sequences into video object planes. The proposed algorithms have three versions: one is based on morphological motion filter and active contour model, and another is based on difference of edges of moving objects and modified Hausdorff distance object tracker. Furthermore, if parts of an object only move occasionally , there is no sufficient motion information using morphological motion filter or difference of edges to identify the whole semantic meaningful object. We propose a new technique which is based on high correlation between edge of frame difference and edge of gray-level image. From collecting enough motion information it can form the binary edge of the whole object and can extract the VOP of the corresponded foreground object.As for the first VOP segmentation algorithm, the initial object is first extracted using morphological motion filter. The filter criterion is the difference between the global dominant motion and local motion. The initial moving object and its binary contour model is obtained after thresholding initial object and removing the noise. Then the binary contour model is tracked and matched to accommodate the change of position and shape of object using active contour model in the following frames. From a series of binary contour models we can extract the video object from the sequences. This algorithm is much suitable for sequences with little motion and stationary or moving background.For the second version VOP segmentation algorithm the moving foreground is identified using absolute difference between edge images first, and its binary edge model is obtained by cannyoperator . In order to accommodate the change of position of object the translation the object has undergone is indicated by the modified Hausdorff object tracker and a two-stage model update method accommodate changes in shape of the object in the following frames. Finally from a series of binary edge models we can extract the video object. This algorithm is much suitable for sequences with fast motion and stationary background.It is difficult to identify the whole semantic meaningful object which suffers from very little moti更多还原