视频序列中基于边缘信息的运动分割方法

【作者】 杜敏松；

【导师】 苏志勋；

【作者基本信息】 大连理工大学 ， 计算数学， 2003， 硕士

【摘要】 近年来，随着MPEG-4压缩标准的制定和越来越多数字视频形式的出现，人们对自动分析视频技术的需求也越来越迫切。在MPEG-4压缩标准中，为了实现视频内容的交互功能，视频序列的每一帧都由视频对象面(VOP)决定，而为了生成这些视频对象面，需要对视频序列中的运动对象进行分割。分析视频的第一步就是把视频序列分割成代表真实世界中不同物体的区域，而且每个区域都遵循着各自对应物体的运动方式。实际上，真正有效的分割技术应该能够提供明确的物体间的划分，以及它们之间相对的深度信息。 运动分割是许多视频分析与应用的必要过程，其目的是把视频序列中的各帧图像按照不同的运动分割成不同的区域，从而分离出图像中的运动对象。本文运用了一种基于边缘信息的运动分割方法，来处理包含前后景两个运动的视频序列。首先用Canny算法提取较为连续的物体边缘，由边缘的法线方向追踪出法向运动，然后对两个运动分别作最大似然法，交替估计出运动和各边缘所属运动的概率，反复迭代直至收敛；再以这些边缘为严格界限，用区域增长法将整幅图像分割成若干区域，最后由于区域标记过程交替估计出正确的运动层次，以确定最后的分割。实验结果表明这种改进的运动跟踪和区域标记方法能够得到很好的分割效果。更多还原

【Abstract】 Following the standard of MPEG-4 and the increasing availability of video in digital form, there is growing demand for methods of automatically analyzing video. In the object-based standard of MPEG-4, in order to support content-based functions, each frame of video should be represented in terms of video object planes (VOP).The first stage is the segmentation of the frames into regions representing different real world objects which have different motion. Exactly, the ideal segmentation should provide a clear partition of objects and the relative depth ordering of each object.Motion segmentation is the necessary stage of most video analysis implements, and it divides frames into different regions according different motion so that those motion objects can be abstracted. This paper presents an improved theoretical framework for motion segmentation based on the motion of tracked edges. Firstly, abstract these edges of objects by Canny algorithm, then tracking edges along normal direction to get normal motion. Secondly, applying Expectation-Maximization algorithm to the two motions alternately to calculate the motions and the responsibility of every edges, repeat the step until it is convergent. Finally segment the sequence into regions by region growing algorithm, and get the final result by region labeling. The practical result has proved the validity of the improved motion tracking and region labeling method.更多还原