Datenbestand vom 20. Januar 2019

Warenkorb Datenschutzhinweis Dissertationsdruck Dissertationsverlag Institutsreihen     Preisrechner

aktualisiert am 20. Januar 2019

ISBN 9783843902236

Euro 84,00 inkl. 7% MwSt


978-3-8439-0223-6, Reihe Informatik

Benjamin Huhle
Acquisition and Reconstruction of 3D Scenes with Range and Image Data

141 Seiten, Dissertation Eberhard-Karls-Universität Tübingen (2011), Softcover, B5

Zusammenfassung / Abstract

We describe a framework for the acquisition and reconstruction of 3D models of real world environments. While commercial products are available featuring accurate 3D scanners and providing tools that allow for the reconstruction of precise and detailed models by relying on extensive user interaction, we focus on the opposite end of the range of acquisition and reconstruction systems reducing costs and efforts for the users. We introduce several sensor setups that consist of comparatively cheap off-the-shelf sensors and concentrate on the integration of different modalities. Utilizing images in addition to range data, we present novel algorithms for several aspects of the reconstruction pipeline. Our denoising technique improves the perceived quality.

Further, the high resolution of the image data can be employed to increase the resolution of the range data. We also present two strategies to include image data in the process of registration of multiple range scans. Features detected in the images are used for a global registration of the range scans, and the relative pose estimates are refined by a geometry-based registration algorithm which we extend to a color-aware version. This leads to robust and accurate results even in cases in which only little texture and few geometrical structures are present.

Based on a novel sensor setup consisting of two 2D laser scanners, we also present a method for the registration of sparse range data which allows to acquire 3D models in a continuous manner.

As we concentrate on the first part of the reconstruction pipeline, the resulting clean and smooth point-based representations of 3D scenes are considered as the result of our systems. To prepare these models for further processing, such as spatial discretization, we present an algorithm that finds the optimal orientation of the model with regard to a global reference system.

Finally, as an alternative to the geometrical reconstruction of 3D scenes, we also present a purely appearance-based approach. In this case, low resolution images from an omnidirectional camera are sufficient to localize a mobile agent based on a model consisting of an interpolated mapping from position to appearance.