Reconstrução geométrica de cenas não estruturadas: uma abordagem monocular com planejamento estocástico

Abstract

This work focuses on the autonomous three-dimensional geometric reconstruction of objects, using a single camera mounted on a mobile robot. Although the geometric reconstruction is a classic Computer Vision problem, most approaches up to date do not deal with the planning aspect, i.e., they do not provide autonomous solutions to determine best camera poses in order to obtain the most complete reconstruction possible. The approach presented here is based solely on images from a single camera and does not depend on any absolute positioning system. As a consequence, this work deals with a class of problems known as Simultaneous Localization and Mapping, or SLAM. In other words, the estimation of the object\\\'s geometry depends on the estimation of current camera poses, which in turn is computed from data that is extracted from the acquired images. The stochastic planning technique developed here requires the continuous determination of the object\\\'s partial geometry. As a matter of fact, partial reconstruction is the single source of information for the planner, which identifies unexplored and under-explored regions and determines the next pose that the camera must adopt in order to continue the exploratory task. This contrasts with several multiple-view geometry algorithms that estimate the entire object\\\'s geometry at once based on a dataset of already available images.