Autonomous Mapping of Unknown Environments Using a UAV
Publicerad
Författare
Typ
Examensarbete för masterexamen
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Automatic object search in a bounded area can be accomplished using cameracarrying
autonomous aerial robots. The system requires several functionalities to
solve the task in a safe and efficient way, including finding a navigation and exploration
strategy, creating a representation of the surrounding environment and
detecting objects visually.
Here we create a modular framework and provide solutions to the different subproblems
in a simulated environment. The navigation and exploration subproblems are
tackled using deep reinforcement learning (DRL). Object and obstacle detection is
approached using methods based on the scale-invariant feature transform and the
pinhole camera model. Information gathered by the system is used to build a 3D
voxel map. We further show that the object detection system is capable of detecting
certain target objects with high recall. The DRL approach is able to achieve navigation
that avoids collisions to a high degree, but the performance of the exploration
policy is suboptimal.
Due to the modular character of the solution further improvements of each subsystems
can easily be developed independently.
Beskrivning
Ämne/nyckelord
Deep reinforcement learning, autonomous exploration and navigation, feature extraction, object detection, voxel map, UAV, modular framework.