Schneider, Johannes: Visual Odometry and Sparse Scene Reconstruction for UAVs with a Multi-Fisheye Camera System. - Bonn, 2020. - , . In: Schriftenreihe / Institut für Geodäsie und Geoinformation, 67.
Online-Ausgabe in bonndoc: http://hdl.handle.net/20.500.11811/8567
@phdthesis{handle:20.500.11811/8567,
author = {{Johannes Schneider}},
title = {Visual Odometry and Sparse Scene Reconstruction for UAVs with a Multi-Fisheye Camera System},
school = {},
year = 2020,
series = {Schriftenreihe / Institut für Geodäsie und Geoinformation},
volume = 67,
note = {Autonomously operating UAVs demand a fast localization for navigation, to actively explore unknown areas and to create maps. For pose estimation, many UAV systems make use of a combination of GPS receivers and inertial sensor units (IMU). However, GPS signal coverage may go down occasionally, especially in the close vicinity of objects, and precise IMUs are too heavy to be carried by lightweight UAVs. This and the high cost of high quality IMU motivate the use of inexpensive vision based sensors for localization using visual odometry or visual SLAM (simultaneous localization and mapping) techniques.
The first contribution of this thesis is a more general approach to bundle adjustment with an extended version of the projective coplanarity equation which enables us to make use of omnidirectional multi-camera systems which may consist of fisheye cameras that can capture a large field of view with one shot. We use ray directions as observations instead of image points which is why our approach does not rely on a specific projection model assuming a central projection. In addition, our approach allows the integration and estimation of points at infinity, which classical bundle adjustments are not capable of. We show that the integration of far or infinitely far points stabilizes the estimation of the rotation angles of the camera poses.
In its second contribution, we employ this approach to bundle adjustment in a highly integrated system for incremental pose estimation and mapping on light-weight UAVs. Based on the image sequences of a multi-camera system our system makes use of tracked feature points to incrementally build a sparse map and incrementally refines this map using the iSAM2 algorithm. Our system is able to optionally integrate GPS information on the level of carrier phase observations even in underconstrained situations, e.g. if only two satellites are visible, for georeferenced pose estimation. This way, we are able to use all available information in underconstrained GPS situations to keep the mapped 3D model accurate and georeferenced.
In its third contribution, we present an approach for re-using existing methods for dense stereo matching with fisheye cameras, which has the advantage that highly optimized existing methods can be applied as a black-box without modifications even with cameras that have field of view of more than 180 deg. We provide a detailed accuracy analysis of the obtained dense stereo results. The accuracy analysis shows the growing uncertainty of observed image points of fisheye cameras due to increasing blur towards the image border. Core of the contribution is a rigorous variance component estimation which allows to estimate the variance of the observed disparities at an image point as a function of the distance of that point to the principal point. We show that this improved stochastic model provides a more realistic prediction of the uncertainty of the triangulated 3D points.},

url = {http://hdl.handle.net/20.500.11811/8567}
}

The following license files are associated with this item:

InCopyright