UBC Theses and Dissertations
An interactive 3D geometric model acquisition and registration system Liu, Yushuang
The acquisition of geometric information from real world objects has now become a major way of modeling complex scenes and environments. Unfortunately, most optical methods for geometric model acquisition require the combination of partial information from different view points, in order to obtain a single, coherent model. This, in turn, requires the registration of partial models into a common coordinate frame, a process that is usually done offline. As a consequence, holes due to undersampling and missing information often cannot be detected until after the registration process. This thesis tries to solve the 3D registration problem using graphics hardware to achieve an interactive speed, so that online next-best-view planning and hole detection during the scanning process can be conducted. The system implemented realizes effective model reconstruction and refinement by interactively using a stereo camera setup and a hardware accelerated depth-map based registration algorithm. The emphasis is put on the hardware accelerated 3D range registration algorithm which can be categorized as a variant of the well known Iterated Closest Point (ICP) algorithm. This work contributes to the field of 3D geometric model acquisition in that it proposes a fast method that leverages the function and structure of modern graphics hardwares. Technically, graphics hardware is used i n two operations. The first is to compute a rigid transformation using modern graphics rendering pipeline. The second is to combine two rendered depth images to compute quickly the absolute difference of z-values of each projected points in the overlap region in the underlying two depth images, where the computation is used to obtain an error metric for a specific candidate transformation during numerical iteration. This error metric is based on depth images which are rendered by point-based rasterization. This system currently performs roughly one registration per second, and is therefore fast enough for on-the-fly evaluation by the user. Given more time, the same method is also capable of producing full geometric models at higher quality.
Item Citations and Data