Use this URL to cite or link to this record in EThOS:
Title: Sparse octree algorithms for scalable dense volumetric tracking and mapping
Author: Vespa, Emanuele
ISNI:       0000 0004 9350 9219
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
This thesis is concerned with the problem of Simultaneous Localisation and Mapping (SLAM), the task of localising an agent within an unknown environment and at the same time building a representation of it. In particular, we tackle the fundamental scalability limitations of dense volumetric SLAM systems. We do so by proposing a highly efficient hierarchical data-structure based on octrees together with a set of algorithms to support the most compute-intensive operations in typical volumetric reconstruction pipelines. We employ our hierarchical representation in a novel dense pipeline based on occupancy probabilities. Crucially, the complete space representation encoded by the octree enables to demonstrate a fully integrated system in which tracking, mapping and occupancy queries can be performed seamlessly on a single coherent representation. While achieving accuracy either at par or better than the current state-of-the-art, we demonstrate run-time performance of at least an order of magnitude better than currently available hierarchical data-structures. Finally, we introduce a novel multi-scale reconstruction system that exploits our octree hierarchy. By adaptively selecting the appropriate scale to match the effective sensor resolution in both integration and rendering, we demonstrate better reconstruction results and tracking accuracy compared to single-resolution grids. Furthermore, we achieve much higher computational performance by propagating information up and down the tree in a lazy fashion, which allow us to reduce the computational load when updating distant surfaces. We have released our software as an open-source library, named supereight, which is freely available for the benefit of the wider community. One of the main advantages of our library is its flexibility. By carefully providing a set of algorithmic abstractions, supereight enables SLAM practitioners to freely experiment with different map representations with no intervention on the back-end library code and crucially, preserving performance. Our work has been adopted by robotics researchers in both academia and industry.
Supervisor: Kelly, Paul H. J. ; Leutenegger, Stefan Sponsor: ARM ; Engineering and Physical Sciences Research Council ; Imperial College London
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral