Reconstruction active par projection de lumière non structurée
memoire » Ph.D.Nicolas Martin
Tags: 3D Vision , Computer vision , Active reconstruction , Coded light , Structured light , Unstructured light , Projector , Camera
Date : 2014-04
Papyrus
http://hdl.handle.net/1866/11197Abstract
This thesis deals with active 3D reconstruction from camera-projector systems. Standard reconstruction methods use coded light patterns that come with their strengths and weaknesses. We introduce unstructured light patterns that feature several improvements compared to the current state of the art. The research presented revolves around three main axes : robustness, precision and comparison of existing unstructured light patterns to existing methods. Unstructured light patterns stand out first and foremost by their robustness to interreflections and depth discontinuities. They are specifically designed to homogenize the indirect lighting generated by their projection on hard to scan surfaces. The downside of these patterns is that matching projected and captured images is not straightforward anymore. A probabilistic correspondence method is formulated to solve this problem efficiently. Another important aspect of reconstruction obtained with unstructured light pat- terns is their ability to recover subpixel correspondences, that is with a precision finer than the pixel level. We present a method to produce long codes using unstructured light. These codes enable us to extract more precise correspondences while requiring less patterns. This contribution makes our method one of the most accurate - yet robust to standard challenges - method of active reconstruction in the domain. Finally, the last part of this thesis adresses the comparison of existing reconstruction methods on several aspects, but mainly on the impact of using less and less patterns on the quality of the reconstruction. While some methods need a fixed number of images, some, like ours, can accommodate fewer patterns in exchange for some quality loss. We devise a simple method to capture an optimal correspondence map that can be used as a groundtruth for comparison purposes. Last, we present several hybrid methods that perform quite well even with few images.
Documents
