| Περίληψη: | As the autonomy and AI technologies are evolving, the challenges in developing com-
plete autonomous missions remain, especially in extreme environments with perceptual
degradation, difficult terrain, high-risk operations and so on. The focus of this Master
Thesis is towards identifying the key modules used for autonomous exploration and map
building with multiple UAVs, and developing important modules that are currently miss-
ing or under-perform in extreme environment scenarios. First, we define a simple decent-
ralized scheme and then move on to discussing the problem of localization and mapping,
as well as our choice of algorithm. During this step, we also present an experimental eval-
uation and a thorough analysis, that highlights the need of a robust global re-localization
algorithm in featureless and GPS-denied environments. Our proposed framework relies
on descriptors extracted through a Siamese Neural Network, that are used for place re-
cognition and yaw regression, yielding a 4DoF transform. Afterwards, we briefly present
the choice of the autonomy framework along with a path planning algorithm. The path
planner provides a trajectory based on the local map and the autonomy framework is re-
sponsible for autonomously following the reference trajectory while avoiding collisions,
identifying objects and so on. Last but not least, we propose a novel framework for ad-
dressing the problem of merging 3D point cloud maps, as it is a module missing from the
current literature. The ability to merge multiple maps during multi-robot exploration is of
high importance as it provides the path planner with a global map and therefore increasing
the efficiency of the algorithm.
|
|---|
