Abstract
A reverse engineering process allows to reconstruct a parametric model from a 3D point cloud or a 3D mesh of an object. The main objective is to retrieve information about the original modeling or fabrication process. The reconstructed model corresponds to a combination of geometric primitives (cylinders, planes, spheres...). These primitives are defined by parameters which are estimated from the 3D input data. Reverse engineering is mainly used in industry, for example in metrology analysis, non-destructive inspection or functional analysis.This thesis, in collaboration with the C4W company, concerns the extraction of geometric primitives from a digitized 3D mesh which represents a manufactured part. This type of mesh often contains many inaccuracies unlike the original CAD model. These inaccuracies can come from both fabrication and digitization processes.First, we try to improve existing techniques to extract robust primitives. Thus, we explore some approaches such as mesh segmentation, shape caracterization based on discrete curvatures and primitive fitting using regression and optimization techniques.The second research axis of this thesis concerns a process called "beautification", which tries to regularize the primitive parameters with respect to geometric relations such as parallelism or concentricity. To do that, we propose an analysis of the primitives, related to standard CAD modeling rules. So, we automatically detect reference elements (coordinate system, planes or lines). In order to guide the model reconstruction, we use these references to regularize the primitives globally.The third research axis concerns the pattern recognition from a set of primitives. So, we search for mechanical features (counterbore, pocket...) and repetitions (linear, circular...), which are often used in CAD modeling of manufactured part. This second analysis, more functionnal, allows to deduce more information related to the modeling and fabrication processes.