Low computational SLAM for an autonomous indoor aerial inspection vehicle

Abstract

The past decade has seen an increase in the capability of small scale Unmanned
\nAerial Vehicle (UAV) systems, made possible through technological advancements
\nin battery, computing and sensor miniaturisation technology. This has opened a new
\nand rapidly growing branch of robotic research and has sparked the imagination of
\nindustry leading to new UAV based services, from the inspection of power-lines to
\nremote police surveillance.
\nMiniaturisation of UAVs have also made them small enough to be practically flown
\nindoors. For example, the inspection of elevated areas in hazardous or damaged
\nstructures where the use of conventional ground-based robots are unsuitable. Sellafield
\nLtd, a nuclear reprocessing facility in the U.K. has many buildings that require
\nfrequent safety inspections. UAV inspections eliminate the current risk to personnel
\nof radiation exposure and other hazards in tall structures where scaffolding or hoists
\nare required.
\nThis project focused on the development of a UAV for the novel application of
\nsemi-autonomously navigating and inspecting these structures without the need for
\npersonnel to enter the building. Development exposed a significant gap in knowledge
\nconcerning indoor localisation, specifically Simultaneous Localisation and Mapping
\n(SLAM) for use on-board UAVs. To lower the on-board processing requirements
\nof SLAM, other UAV research groups have employed techniques such as off-board
\nprocessing, reduced dimensionality or prior knowledge of the structure, techniques
\nnot suitable to this application given the unknown nature of the structures and the
\nrisk of radio-shadows.
\nIn this thesis a novel localisation algorithm, which enables real-time and threedimensional
\nSLAM running solely on-board a computationally constrained UAV in
\nheavily cluttered and unknown environments is proposed. The algorithm, based
\non the Iterative Closest Point (ICP) method utilising approximate nearest neighbour
\nsearches and point-cloud decimation to reduce the processing requirements has
\nsuccessfully been tested in environments similar to that specified by Sellafield Ltd.