【作者】 胡社教；

【导师】 陈宗海；

【作者基本信息】 中国科学技术大学 ， 模式识别与智能系统， 2006， 博士

【摘要】 基于图像绘制是从二十世纪九十年代开始在世界范围内兴起的一门崭新学科，它被广泛应用于虚拟环境的场景构建等领域，是目前信息科学领域的一个研究热点。基于图像绘制的虚拟环境构建是计算机科学和信息科学的交叉和融合，它涉及到计算机图形学、计算机视觉、图象处理、模式识别、人工智能技术等学科的理论和成果。基于图像绘制技术突破了传统基于几何建模和绘制的理论架构，以摄像机拍摄的实景图像为基础，运用图像处理技术对序列实景图像进行拼接，将看上去孤立的图像组织成一幅能够反映全部场景信息的全景图。基于图像绘制的场景构建，避免了基于几何方法的烦琐建模过程，不需要场景的先验知识，直接通过图像变换就能生成具有照片真实感的场景，且绘制速度与场景的复杂度无关，因而特别适合未知环境的场景构建。 具有视觉临场感的机器人遥操作虚拟环境构建是虚拟现实技术在智能机器人领域的重要应用。机器人遥操作系统所面对环境常常都是未知的，因而本文采用基于图像绘制的方法来构建机器人遥操作虚拟环境的场景。为了全面理解基于图像绘制的虚拟环境构建技术，本文首先介绍了传统的基于几何绘制与基于图像绘制的理论基础和发展过程，综述了基于图像绘制技术的研究现状和发展方向。围绕如何把摄像机拍摄的实景序列图像进行拼接，组成一幅能够反映全部场景信息的全景图这一虚拟场景构建技术的核心，提出了几种全景图像的拼接方法，并实现了机器人全景环境漫游。 针对机器人运动时，摄像机在拍摄实景图像过程中可能出现的图像旋转、镜头伸缩和图像传输过程中存在干扰等因素。本文提出了一种基于相位相关的自适应图像拼接算法，本算法首先对原始图像进行边缘提取，得到二值化的边缘化图像，这样可以避免光照条件对图像的影响；对于待拼接的两幅图像存在明显的旋转角度和镜头缩放情况，本文将上述边缘化的图像进行对数极坐标变换，在对数极坐标域中旋转角度和缩放因子就变换成了两个位移量，这样就能方便地使用相位相关法来求得待拼接两幅图像之间的旋转角度和缩放因子，在笛卡儿坐标系中调整两幅图像，使两幅图像之间仅存在位移关系，这样就可以使用基于边缘化图像的相位相关法来实现两幅图像的拼接。 为了提高图像的拼接速度和质量，本文提出了一种基于图像特征角点的二次求精匹配算法，该算法利用角点所具有的旋转不变性和对光照不敏感的特点，首先采用SUSAN算子抽取待拼接两幅图像的特征角点；然后利用快速RANSAC算法，计算图像之间的单应性矩阵：把得到的初始特征角点对作为重叠区域图像中的运动点，引入KLT跟踪算法来精确定位特征角点对的位置，重新计算精确的单应性矩阵；最后对输入图像和参考图像分别采样，注入到一幅更大的图像中，完成图像的拼接融合。 在机器人遥操作虚拟环境构建中，在保证场景信息完整的情况下，使机器人传送的图像数量尽可能的少，并要求全景环境生成实时性好等特点。本文提出了一种基于图像灰度特征矢量和最大匹配度的全景图像拼接方法，该算法选取参考图像的一矩形区域作为匹配块，以矩形块边框的行列像素灰度差作为两个图像灰度特征矢量，在待匹配区域寻找相应的匹配块更多还原

【Abstract】 Image-based rendering is a new subject that has been developed since 20 century 90 decade in the world, and has been applied extensively in building of virtual environment. It is a present research hot point of information scientific field. Image-based rendering virtual environment building is the intersect and fusion of computer science and information science, it is concerned with theory and accomplishment of computer graphics, computer vision, image processing, pattern recognition and artificial intelligence subjects. It has broken through the traditional theoretical configuration of geometry-based modeling and rendering, used the image processing technology to mosaic a isolated sequence real scene image and formed a panorama that can reflect completely scene information. Image-based rendering scene building technology avoid the overelaborate of geometry-based modeling method, generate scene of photograph real sense by direct image alternate and don’t need scene examine knowledge, rendering speed is unconcerned the complex degree of scene, suit the scene building of unknown environment especially.The scene building of having vision telepresence robot teleoperation virtual environment is virtual reality technical important application in intelligent robot field. Robot teleoperation system face an unknown environment usually, so image-based rendering is used to build robot teleoperation virtual environment. To understand image-based rendering virtual environment building technology overall, this paper has first introduced the theoretical basic and development of traditional geometry-based rendering and image-based rendering, then summarized research present situation and develop direction of image-based rendering technology. Revolve around how to used the image processing technology to mosaic a isolated sequence real scene image and formed a panorama that can reflect completely scene information, some kinds of panoramic mosaic mothd has put forward, and realized the wander in robot panoramic environment.According to robot sport, camcorder image may appear spin, camera lens flexible and image transmission course in existence disturb. This paper have put forward a kind of based on phase correlate adaptive image mosaic algorithm that first carries out the edge of original image to get 2 values the edge, so avoid illumination condition the influence to image. The edge image will be alternated to logarithm polar coordinates, in logarithm polar coordinates the revolving angle and camera lens flexible are alternated two displacements , so the phase correlate law can be used conveniently to get the revolving angle and camera lens flexible. Then in descartes coordinates the two images of joining together is adjusted, used the edge image phase correlate law to make storing only between two images reign and joining together that realizes two images mosaicing.To raise speed and precision of mosaic panorama, a kind of image-based feature corner-point twice match algorithm is put forward, the algorithm utlize the revolving consistency and not sensitive illumination of corner-point. At first the SUSAN operator is used to take out the feature corner-point of the two adjacent images. Then using the fast RANSAC algorithm to calculate the homography matrix, and feature corner-point in overlapping regional image is regard as sport points, KLT trail algorithm is used in the accurate location of feature corner-point, homography matrix is calculated accurately. Finally, the differentiate sampling for input image and reference image, pour into a larger image, and realize the mosaic and fusion of panorama.In building robot teleoperation virtual environment, when complete scene information is更多还原