Dense real-time 3D reconstruction from multiple images

Ling, L 2013, Dense real-time 3D reconstruction from multiple images, Doctor of Philosophy (PhD), Electrical and Computer Engineering, RMIT University.


Document type: Thesis
Collection: Theses

Attached Files
Name Description MIMEType Size
Ling.pdf Thesis application/pdf 4.65MB
Title Dense real-time 3D reconstruction from multiple images
Author(s) Ling, L
Year 2013
Abstract The rapid increase in computer graphics and acquisition technologies has led to the widespread use of 3D models. Techniques for 3D reconstruction from multiple views aim to recover the structure of a scene and the position and orientation (motion) of the camera using only the geometrical constraints in 2D images. This problem, known as Structure from Motion (SfM) has been the focus of a great deal of research effort in recent years; however, the automatic, dense, real-time and accurate reconstruction of a scene is still a major research challenge. This thesis presents work that targets the development of efficient algorithms to produce high quality and accurate reconstructions, introducing new computer vision techniques for camera motion calibration, dense SfM reconstruction and dense real-time 3D reconstruction. In SfM, a second challenge is to build an effective reconstruction framework that provides dense and high quality surface modelling.

This thesis develops a complete, automatic and flexible system with a simple user-interface of `raw images to 3D surface representation'. As part of the proposed image reconstruction approach, this thesis introduces an accurate and reliable region-growing algorithm to propagate the dense matching points from the sparse key points among all stereo pairs. This dense 3D reconstruction proposal addresses the deficiencies of existing SfM systems built on sparsely distributed 3D point clouds which are insufficient for reconstructing a complete 3D model of a scene. The existing SfM reconstruction methods perform a bundle adjustment optimization of the global geometry in order to obtain an accurate model. Such an optimization is very computational expensive and cannot be implemented in a real-time application. Extended Kalman Filter (EKF) Simultaneous Localization and Mapping (SLAM) considers the problem of concurrently estimating in real-time the structure of the surrounding world, perceived by moving sensors (cameras), simultaneously localizing in it. However, standard EKF-SLAM techniques are susceptible to errors introduced during the state prediction and measurement prediction linearization. 
Degree Doctor of Philosophy (PhD)
Institution RMIT University
School, Department or Centre Electrical and Computer Engineering
Keyword(s) 3D Reconstruction
Cheiralilty
SfM
SLAM
EKF
Versions
Version Filter Type
Access Statistics: 404 Abstract Views, 1516 File Downloads  -  Detailed Statistics
Created: Wed, 28 May 2014, 12:29:17 EST by Denise Paciocco
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us