Panorama Generation for Stereoscopic Visualization of Large-Scale Scenes

In this thesis, we address the problem of modeling and stereoscopically visualizing large-scale scenes captured with a single moving camera. In many applications that image large-scale scenes the critical information desired is the 3D spatial information of stationary objects and movers within the scene. Stereo panoramas, like regular panoramas, provide a wide field-of-view that can represent the entire scene, with the stereo panoramas additionally representing the motion parallax and allowing for 3D visualization and reconstruction of the scene. The primary issue with stereo panorama construction methods is that they are constrained for a particular camera motion model; typically the camera is constrained to move along a linear or circular path. Here we present a method for constructing stereo panoramas for general camera motion, and we develop a (1) Unified Stereo Mosaic Framework that handles general camera motion models. To construct stereo panoramas for general motion we created a new (2) Stereo Mosaic Layering algorithm that speeds up panorama construction enabling real-time applications. In large-scale scene applications it is often the case that the scene will be imaged persistently by passing over the same path multiple times or two or more sensors of different modalities will pass over the the same scene. To address these issues we developed methods for (3) Multi-Run and Multi-Modal Mosaic Alignment. Finally, we developed an (4) Intelligent Stereo Visualization that allows a viewer to interact and stereoscopically view the stereo panoramas developed from general motion