Pose Imagery and Automated Three - DimensionalModeling of Urban

Three-dimensional (3-D) modeling of urban environments has numerous applications, including virtual environments, urban planning, and physical simulation. Constructing 3-D models from photographs (images) is thus an important area of research in computer vision, and increasingly, computer graphics. However, despite many years of research, a system that automatically recovers realistic 3-D models remains elusive; most practical systems require signiicant human input. Unlike automatic algorithms, human-assisted systems are not scalable, both in terms of the number of images processed and the complexity of the generated 3-D model. This thesis describes novel techniques to automatically extract textured 3-D models of urban environments from pose imagery, i.e., images annotated with camera position and orientation in a single global coordinate system. Physical instruments (e.g., surveying, Global Positioning System (GPS), inertial sensors, etc.) are used to provide accurate initial pose estimates to the proposed algorithms. As these estimates are not perfect, I rst describe two optimization techniques that reene pose estimates using information present in the images: spherical mosaicing recovers relative rotations between images taken from a single position, and mosaic registration accurately locates mosaics in a global coordinate system. Next, I describe an algorithm that extracts vertical facades from mosaics annotated with accurate pose. The algorithm employs horizontal line segments to detect likely facade orientations and locates these facades using a space-sweep technique. Textures are robustly computed for the facades by combining information from several mosaics using median statistics. I present results for a large pose image dataset (consisting of about four thousand images taken from eighty-one positions) of an urban ooce complex. These techniques were successful in recovering all signiicant vertical facades in the complex, as well as several neighboring facades. 3 4 Acknowledgements I rst would like to thank Prof. Seth Teller for being an absolutely wonderful thesis supervisor through the entire process. He started me out with great ideas and problems, provided good feedback and suggestions during the research, and made sure that the nal output is of high quality. As well as teaching me graphics, he showed me that good research can be fun. I thank him for making my time at MIT a very rewarding experience. Thanks to Prof. Tomm as Lozano-P erez and to Prof. Leonard McMillan for serving as readers on my committee and providing very useful comments on the thesis draft. Thanks also to Prof. Julie Dorsey for her interest in my research. I am grateful to …