Acquisition of Depth Information form Motion Images Taken in Natural Scene and Their Synthesis.

本课题研究了从图像中获取深度信息的几种方法，并提出了一种基于深度信息的图像合成方法。首先，研究了从多视点图像中提取极极几何深度信息的两种方法：一种是立体图像，另一种是多视点系统的特殊情况，即单目图像拍摄有镜子的场景。在立体视觉方法中，通过相关技术计算显著边缘上的视差，将行线上两个边缘之间的点与边缘的视差进行线性插值，并对插值后的视差进行校正，使其与垂直扫描得到的视差一致。实验结果表明，该方法简单地逼近了一个深度几乎均匀的区域，并将前景区域与背景区域分离开来。另一种使用镜像的方法是基于仿射变换。在该方法中，当给定桌子上物体的四个点及其对应的反射在镜子上的点时，就可以确定桌子上另一个三维物体投影在像面上的二维点。这意味着一个真实视频的CG动画和一个运动中的三维物体可以从一个特殊的单目图像生成。其次，研究了基于三维几何的图像三维物体模型重建方法，用于图像合成。提出了一种基于模型的人体运动匹配方法。通过将模型与由四幅不同视点的图像构成的体素数据进行匹配，证明了识别每个姿态的运动参数是可能的。第三，研究了基于轮廓的目标跟踪提取深度信息。目标跟踪可以在场景中分离前景区域和背景区域。在此方法中，提出了一种新的基于Hopfield神经网络的区域分割方法。将边界像素的某些属性嵌入到网络的目标函数中，得到使该函数最小的分割结果。为了实现快速、稳定、全局的区域分割，采用了金字塔图像。实验结果表明，该方法对区域提取和运动跟踪是有效的。提出了一种新的无运动不相容的图像合成方法。该方法在第一帧中指定待叠加目标的轮廓后，采用层次匹配算法和主动轮廓法在目标图像序列中进行半自动跟踪。为了合成无运动不相容的两幅图像序列，根据轮廓跟踪过程中提取的运动参数，自动确定待叠加对象的位置和放大比。针对运动模糊问题，提出了一种新的边界颜色混合软键。少

In this project, several methods for acquiring depth information from images were investigated and an image synthesis method using the information was proposed.Firstly, two approaches, in which epipolar geometry derived from multi-view images was analyzed to extract depth information, were investigated : one is stereo images and the other is a special case of multi-view system, that is, a monocular image taking a scene in which there exists a mirror. In the approach of stereo vision, disparities on significant edges were calculated by a correlation technology, and a point between two edges in a row line is linearly interpolated its disparity from those of the edges, and the interpolated disparity is corrected so as to be consistent with that obtained by vertical scan. Experimental results showed that this method simply approximate a region with almost homogeneous depth and also separate foreground regions from backgrounds ones. The other approach using a mirror is based on affine trans … More formation. In this method, when four points of an object on desk and their corresponding points reflected on the mirror are given, it is possible to determine 2-D points projected on the image plane of another 3-D object on the desk. This means that CG animation of a real video and a 3-D object in motion can be generated from a special monocular image.Secondly, 3-D object model reconstructions from images based on 3-D geometry were investigated for image synthesis. A model based volume matching method to extract motion of human being was proposed. It was proved to be possible to identify motion parameters for each pose by matching models with voxel data constructed by four images from different viewpoints.Thirdly, object tracking by contours to extract depth information were investigated. The object tracking enables to separate foreground regions from background ones in a scene. In this approach, a new region segmentation method based on the Hopfield NN was proposed. Some properties of boundary pixels were embedded in an objective function of the network and the segmentation which minimizes this function was obtained. In order to realize a fast, stable and global region segmentation, pyramid images were used. Experimental results showed that this method is effective for region extraction and motion tracking.A new image synthesis method without motion incompatibility was also proposed. In this method, when the contours of objects to be superimposed are specified once in the first frame, they are semi-automatically tracked in object image sequence by a hierarchical matching algorithm and an active contour method. In order to synthesize two image sequences without motion incompatibility, position and magnification ratio of an object to be superimposed are determined automatically based on its motion parameters extracted in the process of contour tracking. A new soft-key for blending boundary colors was also proposed to cope with motion blurs. Less