Volumetric object reconstruction by means of photogrammetry

v Title of Dissertation: Volumetric Object Reconstruction By Means of Photogrammetry ABSTRACT: In many fields such as virtual reality, science, medicine and entertainment, an automated generation of 3D computer models is required. In this thesis the problem of creating photorealistic 3D models from still image sequences is addressed and a volumetric object reconstruction technique based on voxel models is presented. One of the goals of this study is to keep the system requirements as low as possible so that the algorithm can be used in many application areas. A standard CCD video camera is used which can capture still images in high resolution. A circular camera setup was done, which was actually accomplished by rotating the object itself, while the camera position was stationary. After calibrating the camera, the orientation parameters of the acquired images are calculated with a bundle block adjustment. A popular computer vision technique is used in order to achieve an approximate model since this technique is fast and easy to reconstruct 3D models of real-world objects. In order to refine the model and get more precise results, conventional photogrammetric techniques are often referred. The object is reconstructed in several steps. The algorithm presented begins by initializing a volume that encloses the 3D objects to be reconstructed. A first approximation of the model is acquired by shape from silhouette which delivers the objects visual hull. Unfortunately, this method does not recover concavities on the object. In order to refine the representation, several tools are described. An important factor is the visibility information of a voxel in a specific image. It can be occluded for several reasons. A fast method is presented to recover this information with a line tracing algorithm. Furthermore, a quality measure of the visibility is introduced, by using the surface normal vector of a voxel in combination with the viewing direction of the image. These tools will help to generate the upgraded model more accurately. The objects natural texture will serve as criterion to exclude non-surface voxels. Since only surface voxels will be projected into corresponding image points in an oriented bundle, we can make use of a similarity value to distinguish surface from non-surface voxels. Several algorithms to evaluate similarity are introduced and compared. With the help of the tools introduced, the concave areas on the object’s surface are recovered successfully and the final model represents the object correctly. In many fields such as virtual reality, science, medicine and entertainment, an automated generation of 3D computer models is required. In this thesis the problem of creating photorealistic 3D models from still image sequences is addressed and a volumetric object reconstruction technique based on voxel models is presented. One of the goals of this study is to keep the system requirements as low as possible so that the algorithm can be used in many application areas. A standard CCD video camera is used which can capture still images in high resolution. A circular camera setup was done, which was actually accomplished by rotating the object itself, while the camera position was stationary. After calibrating the camera, the orientation parameters of the acquired images are calculated with a bundle block adjustment. A popular computer vision technique is used in order to achieve an approximate model since this technique is fast and easy to reconstruct 3D models of real-world objects. In order to refine the model and get more precise results, conventional photogrammetric techniques are often referred. The object is reconstructed in several steps. The algorithm presented begins by initializing a volume that encloses the 3D objects to be reconstructed. A first approximation of the model is acquired by shape from silhouette which delivers the objects visual hull. Unfortunately, this method does not recover concavities on the object. In order to refine the representation, several tools are described. An important factor is the visibility information of a voxel in a specific image. It can be occluded for several reasons. A fast method is presented to recover this information with a line tracing algorithm. Furthermore, a quality measure of the visibility is introduced, by using the surface normal vector of a voxel in combination with the viewing direction of the image. These tools will help to generate the upgraded model more accurately. The objects natural texture will serve as criterion to exclude non-surface voxels. Since only surface voxels will be projected into corresponding image points in an oriented bundle, we can make use of a similarity value to distinguish surface from non-surface voxels. Several algorithms to evaluate similarity are introduced and compared. With the help of the tools introduced, the concave areas on the object’s surface are recovered successfully and the final model represents the object correctly.