@mastersthesis {pub2798,
	title = {Robust Visual Features for Long-term Stable Outdoor Self-localization of Autonomous Service Robots},
	author = {Annika Besetzny},
	year = {2014},
	month = {October},
	abstract = {The aim of this thesis is to identify robust visual features in an unknown scene, which
an autonomous vision-based service robot can use to build a long-term stable map and
to localize itself in an outdoor environment still after a long period of time.
With the detector-descriptor combination FAST and BRIEF and the OpenCV bruteforce
matcher, stable and unstable features are identified on long-term time lapse movies
of different environments. Compared to conventional methods, in this thesis an additional
filter step with a selection criterion is used between the feature detection and
description. In that filter step keypoints are identified that are more likely robust over
longer time periods. The additional selection criterion, a combination of a Gabor filterbank
and a hue histogram, is applied on the stable and unstable keypoints. Those
feature vectors are used to train a support vector machine by supervised learning. The
generalization ability of the selection criterion is demonstrated by testing the trained
classifier on a different test data set. It successfully identifies the robust keypoints.
With the ability to find robust features on a changing terrain, an autonomous service
robot is able to localize itself when weather or lighting conditions change or even in a
different season.
},
	publisher = {Frankfurt University of Applied sciences},
	booktitle = {Frankfurt University of Applied sciences}
}