Titolo della tesi: Stereo Visual SLAM and Place Recognition with Binary Feature Descriptors
Simultaneous Localization and Mapping (SLAM) has been and still is the cornerstone for
autonomy of past and current mobile robots. In most applications, industrial or domestic, the
estimated state comprises position and imminent surroundings. Independent of the manifold
of different systems, reliability is crucial for effective operation and collision avoidance.
In research, the driving algorithms constituting SLAM state estimation are often highly
complex and entailed to a particular use case. In this thesis we carefully examine the state-
of-the-art of such systems and propose novel approaches and improvements accompanied by
insights in a systematic and straightforward algorithmic architecture. More specifically, we
present the results of our research on SLAM and Visual Place Recognition (VPR) utilizing
binary feature descriptors.
First, we contribute to the Robotics research community by releasing a novel, complete
stereo and depth SLAM system. Our system stands out from the state-of-the-art in its
simplicity, modularity, clean object-oriented design and processing speed. Thanks to the
object-oriented design, the two sensor modalities are handled with minimal changes in the
pipeline. We show that our lightweight approach performs head-to-head with established
state-of-the-art systems in a series of evaluations on standard SLAM benchmark datasets.
Second, we complement our first contribution with the publication of an novel VPR
approach for similarity search. In contrast to most existing similarity search solutions
employed in SLAM, our approach is tailored to the binary feature-based case. This enables
us to process images up to two magnitudes faster than compared systems while maintaining
a competitive accuracy. An extensive benchmark evaluation on SLAM datasets demonstrates
the soundness and applicability of our approach to the SLAM use case.
Third, we propose a generic method for improving accuracy and robustness of binary
descriptor-based VPR solutions. Our method is independent of the descriptor type and
compatible with common similarity search approaches. We validate our method in a rigorous
experimental evaluation on a multitude of public benchmark datasets and state-of-the-art
reference approaches.
All of our contributions are available as open source. We believe that this will foster
collaboration and enable straightforward repeatability of our experiments. Every repository
is accompanied by a wiki that both explains the usage of our software and provides additional
experimental results to the evaluations presented this document.