Open access
Author
Date
2019-10Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
Visual (self)localization enables Autonomous Ground Vehicles (AGVs) to assess their
position and orientation within an environment with up to centimeter level accuracy,
using only cost-effective camera sensors. Especially for high precision maneuvering
in GNSS-denied environments, using cameras for localization may be the best
suited option for budget- or weight constrained platforms. However, particularly in
outdoor environments, camera images are subject to various forms of appearance
change. This renders it challenging to reliably localize a vehicle against a map
previously built from sensor data recorded under different appearance conditions.
A powerful approach to deal with these appearance changes is to enhance the map
with visual data from several recordings, each collected under different appearance
conditions. The amount of data generated following this approach, however, scales
with the number of recordings collected over time, and thus unveils a need for
smart algorithms managing this data and ensuring efficient use of computation,
storage and network bandwidth resources. The contributions of this thesis are
centered around the research questions addressing this need for a resource-efficient
and reliable visual localization system for AGVs in outdoor environments.
In Part A, we propose an algorithm to dynamically select small amounts of
map data matching the current appearance condition, thereby lowering network
bandwidth consumption, and reducing computational demands on the vehicle
platforms. We show that exploiting co-observability statistics allows for performing
this appearance-based map data selection in a highly effective manner, without the
need to explicitly model or enumerate the different appearance conditions.
Part B is devoted to the development of a practical map management process
for a visual localization system targeted at long-term use. Our experiments have
revealed that multi-session maps converge to a relatively stable state after several
months of collecting recordings under varying appearance conditions. Furthermore,
through a tight integration of appearance-based map data selection with offline map
summerization, a completely scalable visual localization and mapping framework is
reached that can be used for indefinite periods of time.
In Part C, we present the visual localization system developed within the
UP-Drive project 1 for autonomous cars in urban outdoor environments. Thereby,
a special focus has been placed on robustness against outdoor and long-term ap-
pearance change, and on a careful evaluation of the localization accuracy. We
demonstrate that reliable and accurate visual localization is feasible in structured
outdoor environments, even over long time spans, across vastly different seasonal, weather, and lighting conditions including at night-time, and with local point
features with binary descriptors on a CPU-only computer architecture. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000372042Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZurichOrganisational unit
03737 - Siegwart, Roland Y. / Siegwart, Roland Y.
More
Show all metadata
ETH Bibliography
yes
Altmetrics