Efficient 3d scene modeling and mosaicing
- Autores: Tudor Nicosevici
- Directores de la Tesis: Rafael Garcia Campos (dir. tes.)
- Lectura: En la Universitat de Girona ( España ) en 2009
- Idioma: inglés
- Tribunal Calificador de la Tesis: Jorge Manuel Miranda Dias (presid.), Xavier Cufí i Soler (secret.), Xavier Lladó Bardera (voc.), Pierre Drap (voc.), Antoni Grau Saldes (voc.)
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
Scene modeling has a key role in applications ranging from visual mapping to augmented reality, This thesis presents an end-to-end solution for creating accurate 3D textured models using monocular video sequences, with contributions at different levels.
First, we discuss a method developed within the framework of sequential Structure from Motion, where a 3D model of the environment is maintained and updated as new visual information becomes available. The camera pose is recovered by directly associating the 3D scene model with local image observations, using a dual registration approach. Compared to the standard Structure from Motion techniques, this approach decreases the error accumulation while increasing the robustness to scene occlusions and feature association failures, allowing 3D reconstructions for any type of scene.
We also develop an online approach for measuring similarities between images. In this way, images corresponding to the same scene region can be associated, allowing the reduction of drift and position uncertainties for mapping and navigation. Inspired from content-based image retrieval, the proposed approach uses visual vocabularies to represent images as occurrences of visual words. The technique is entirely sequential and automatic, making it suitable for online applications, such as robot navigation and mapping: (\emph{i}) the vocabularies are built and updated online, during image acquisition, in order to efficiently represent the visual information present in the scene, and (\emph{ii}) the vocabulary building and image indexing processes do not require any user intervention.
Lastly,motivated by the need to map large areas, we propose an online 3D model simplification algorithm. The simplification process uses plane-parallax to estimate the geometry of the scene, eliminating the necessity of explicit scene shape information. Such an approach offers two main advantages: (\emph{i}) it is suited for online applications, where it can run parallel with the 3D reconstruction process, and (\emph{ii}) as it does not require having the full model prior to the simplification, the algorithm allows mapping of larger, more complex scenes.
We discuss the efficiency of the proposals and compare them with other state of the art approaches, using a series of challenging datasets both in underwater and outdoor scenarios.
