Use this URL to cite or link to this record in EThOS:
Title: Loop Closing Detection in SLAM Using Scene Appearance
Author: Ho, Kin Leong
ISNI:       0000 0001 3579 3168
Awarding Body: University of Oxford
Current Institution: University of Oxford
Date of Award: 2007
Availability of Full Text:
Full text unavailable from EThOS.
Please contact the current institution’s library for further details.
This thesis is concerned with the detection of loop closing in a Simultaneous Localisation arid Mapping (SLAM) application. The loop closing detection problem asks how a robot can 'recognise' it has returned to a previously visited location after completing a long circuitous path. Many SLAM implementations look to internal map and pose estimates to make decisions about whether a robot has closed a loop. Approaches that rely on these estimates are generally unreliable in detec:ting loop closure when true and estimated robot poses diverge greatly. The aim of this thesis is to produce a loop closing detection algorithm that is independent of pose estimates, sensor modality and estimation techniques, and works across a spectrum of workspaces. A key competency required is appearance-based place recognition. In order to achieve this goal, some significant issues pertinent to place recognition, namely perceptual variability and aliasing, have to be resolved. Viewpoint invariant descriptors are derived from observations used to represent local scenes. An efficient retrieval system coupled with indexing techniques allows for rapid comparison between observations based on a similarity function. Similarity relationships between local scenes are then encoded within a similarity matrix. The loop closing problem is then addressed as a sequence detection problem within a similarity matrix. Exploiting the phenomenon that loop closing events occur as off-diagonals within a similarity matrix, a sequence detection algorithm is developed to extract such sequences. Instead of finding matching pair of observations, matching sequences are detected so as to e>.-ploit the topological relationships between scenes to reduce false positives. To further tackle the perceptual aliasing problem, spectral decomposition of a similarity matrix is carried out. The effects of repetitive and ambiguous artefacts found within an environment are removed through rank reduction based on an entropy maximisation criterion. Sequence detection is achieved in these rank reduced matrices. A principled manner to determine the probability of sequence occurring randomly allows the evaluation of the significance of such sequences before loop closing is triggered. The practical implementation of the loop closing technique is demonstrated in a variety of challenging scenarios and experimental settings.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available