Reconstruction 3D sous GPU à partir d'un ensemble réduit d'images ordonnées
Document typeMaster thesis
Rights accessRestricted access - confidentiality agreement
This document is the report of the final master thesis. This project has been done in the context of a six-month stage convention between THALES Security Solutions & Services and the Barcelona Informatics Faculty (FIB). We present a GPU accelerated implementation of a global 3D scene reconstruction chain from uncalibrated images. The lack of the actual systems is that they¿re not real-time. GPU programming allows producing new algorithms for a robust 3D reconstruction with real-time computation. In this document we give a set of basic theory concepts in this area and also a complete guide to implement the 3D reconstruction algorithms using the newest graphics hardware and features. Here we will explain deeply the context and the project itself, from the conception, with its theory and concepts, to the design and implementation. In addition we show a non-detailed version of the code that will help the comprehension of ideas explained in the implementation sections. Some introduction to the state of art of the project and to the graphics hardware concepts is also given. We strongly recommend reading these chapters before starting the architecture section, if not, there will be lots of misunderstood concepts. We will assume some previous knowledge in computation theory and geometrics. Anyway, we added a term glossary that contains all the keywords with a short explanation. We will also discuss the results of this new implementation and give further ideas to improvements. By now, we got a very important gain of speed, compared to the previous conceived CPU version. While this new system does not actually satisfy the real-time constraint, we are able to reconstruct 3D buildings in less than one minute.