Study of parallel techniques applied to surface reconstruction from unorganized and unoriented point clouds

Abstract

Nowadays, digital representations of real objects are becoming bigger as scanning processes are more accurate, so the time required for the reconstruction of the scanned models is also increasing. This thesis studies the application of parallel techniques in the surface reconstruction problem, in order to improve the processing time required to obtain the final mesh. It is shown how local interpolating triangulations are suitable for global reconstruction, at the time that it is possible to take advantage of the independent nature of these triangulations to design highly effcient parallel methods. A parallel surface reconstruction method is presented, based on local Delaunay triangulations. The input points do not present any additional information, such as normals, nor any known structure. This method has been designed to be GPU friendly, and two implementations are presented. To deal the inherent problems related to interpolating techniques (such as noise, outliers and non-uniform distribution of points), a consolidation process is studied and a parallel points projection operator is presented, as well as its implementation in the GPU. This operator is combined with the local triangulation method to obtain a better reconstruction. This work also studies the possibility of using dynamic reconstruction techniques in a parallel fashion. The method proposed looks for a better interpretation and recovery of the shape and topology of the target model.