Abstract:
An increasing number of modern Computer Vision applications rely on the usage of depth data, allowing the accomplishment of several tasks otherwise being complex or impractical. As a consequence, a large amount of 3D acquisition methods have been proposed in literature, each one designed for a specific objective. In particular, structured-light approaches employ a projector to cast additional information over the scene to compute correspondences between two views and obtain a dense reconstruction.
This thesis presents two main contributions in the field of 3D reconstruction: first, some improvements over a state-of-the-art structured light technique called fringe projection are presented. We start from a theoretical characterization of the acquired signal and formulate a novel phase unwrapping technique offering high resilience to noise and outliers. Additionally, we propose a code recovering technique and a phase correcting method to further improve the final reconstruction accuracy. The second main contribution involves some real-world applications exploiting 3D reconstruction techniques in heterogeneous fields. In particular, we present a cylinder extraction technique designed to work in industrial settings where non-oriented, noisy point clouds are acquired from the scene and a structured-light approach for micrometric surface reconstruction. Another application involves the employment of the described techniques in the field of cultural heritage, where the final outcome yields to a reverse engineering process in addiction to the main preservation purpose.
