【作者】 令狐永芳；

【导师】 王士同；

【作者基本信息】 江南大学 ， 计算机应用技术， 2007， 硕士

【摘要】 随着计算机技术的不断进步,计算机视觉的研究对象已经从单幅静止图象的研究转到对运动图像序列的研究。运动分割是把序列图像划分为在语义上具有不同意义区域,进而分割出运动物体的过程,它是许多运动图像分析应用中必不可少的初始处理阶段。但是要提出准确的、高效的运动分割方案仍然是一个具有挑战性的任务。本文首先叙述了运动对象分割技术的产生和发展,讨论了运动图像分割技术的现状。阐述了该领域的基本研究方法,并对这些方案进行了比较系统的分类。然后探讨了一种基于马尔可夫随机场(MRF)模型的运动目标自动分割算法。该算法采用高斯混合分布描述视频序列的差分图像,对标准MAP算法进行了改进,使用快速方法计算后验边缘。先对视频处理对象进行初始分割,获取初始运动数目以及相应的运动模型的初始参数,然后通过参数估计,不断更新模型参数,之后通过把每个运动区域和运动模型相关联,来同时估计多个运动区域,最终达到分割的目的。经实验验证,本文所提的方法是有效的。最后,在前面工作的基础上,将空间域的分割结果作为图像的观察场,时间域分割结果作为初始标记场,然后利用模型的约束条件将二者结合起来,得到该帧最后的分割标记场,达到分割的目的。通过实验结果证明,本文所提的方法对运动目标分割具有较好的分割效果。更多还原

【Abstract】 With the development of computer technology, the object of study has been transferred from single static image to motorial image sequences. Motion segmentation which is an initial and necessary stage in many video analysis applications is the process of obtaining moving objects by dividing video frames into regions that have different motions. However, it is still a challenge to provide an accurate and efficient motion segmentation method.This paper first introduces the production and development of motorial object segmentation technology and its present situation. Then discuss the basic research method of this field and give the systematic classification to these methods. The following part of the paper, a novel video motion object automatic segmentation algorithm based on Markov random field has been studied in detail. This algorithm use Gaussian mixture distributions to describe the different images of video sequence and make some improvement to the standard MAP algorithm by using the fast method to compute the posterior marginal. First of all, initial segmentation of video is used to obtain the number of initial motions and the corresponding initial parameters of the motion model. And then it connects every motion area and motion model to estimate multi-motion area synchronously by updating the model parameters, consequently achieve the aim of segmentation. The experiments show that the method is effective.In addition, based on above work, the spatial segmentation is provided as an observed field of the image and the field is initialized as the temporal segmentation result. Then they are connected by the model restraint condition to obtain the final labels. With the experiment results, the above algorithm has good performance to motion object segmentation.更多还原