Rekonstruktion und Verarbeitung von Objekten und Szenen aus Kamerabildern

Highly detailed models are commonly used in computer games, animation films and other interactive rendering applications. In most cases the models correspond to a real object, which can be scanned using a laser scanner. Because modern laser scanners are expensive and unwieldy, alternatives are needed, which can generate a virtual duplicate in an economic and efficient way. In addition to the methods for the reconstruction of three-dimensional objects, intuitive editing techniques are also required. In this context multi resolution modeling algorithms are very popular to perform the editing process of a model in real time and thus simulate complex motion sequences, for instance. This dissertation presents a novel solution for the reconstruction of three-dimensional objects and scenes using two-dimensional images and introduces new techniques for the real-time editing of a given polygon mesh. By using reciprocal pairs of images the correspondences between the individual pixels can be uncovered and thus a depth analysis can be realized. The resulting depth values are then stored in a depth map, which ultimately is used to generate a three-dimensional triangulated mesh. During the implementation of the procedure the parallelization of the individual steps of the calculations represents the most important part. For the modeling of a 3d object a high-quality parallel simplification algorithm was proposed, which is able to create multiple resolutions of a complex triangle model in real-time. Based on this simplification procedure a parallel multi resolution modeling algorithm was finally realized, which can be used to modify the overall shape or small scale details of a 3d object at different levels. The completed modifications are propagated to all resolutions by preserving all details of the original surface. All required data structures were generated by a massively parallel simplification algorithm within a few seconds.