E-Thesis 201 views 64 downloads
Deep Visual Place Recognition for Shoreline Navigation / LUKE THOMAS
Swansea University Author: LUKE THOMAS
-
PDF | E-Thesis – open access
Copyright: The author, Luke Thomas, 2025 Distributed under the terms of a Creative Commons Attribution 4.0 License (CC BY 4.0)
Download (83.35MB)
DOI (Published version): 10.23889/SUThesis.70032
Abstract
The use of visual place recognition (VPR) to identify approximate geographical locations from land-based imagery has seen great success in recent history, generally methods of performing VPR work by converting images into a set of representative feature descriptors which can then be compared mathema...
| Published: |
Swansea University, Wales, UK
2025
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Roach, M., Edwards, M., Rahat, A., and Capsey, A. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa70032 |
| Abstract: |
The use of visual place recognition (VPR) to identify approximate geographical locations from land-based imagery has seen great success in recent history, generally methods of performing VPR work by converting images into a set of representative feature descriptors which can then be compared mathematically to determine a closest match. Originally, these descriptors were generated using hand-crafted methods such as Scale Invariant Feature Transform (SIFT),however in more recent years we have seen the development of Deep VPR, extending the methodology to use outputs from Convolutional Neural Networks (CNN) as the basis for feature descriptors instead, an idea that has lead to the development of multiple state-of-the-art methods in the last decade. However, almost all of these works are focused on land-based imagery exclusively which makes sense given that one of the biggest motivators behind the development of Deep VPR is assisting in the navigation of autonomous cars and robots. Our work seeks to test the viability of Deep VPR for Shoreline Imagery for the purpose of sea vessel navigation, where images still contain nearby land features from the visible shoreline but are taken from a vessel out on the water thus introducing a drastically different perspective from the types of images most state-of-the-art Deep VPR models have been trained and evaluated on. In this thesis we provide a new in-house dataset containing images generated form several recordings of travels across the Plymouth Sound, UK over multiple days during March and April 2022 provided by a mounted camera system placed on the IBM/Promare Mayflower Autonomous Ship. This dataset forms our benchmark for evaluating Deep VPR performance on Shoreline Imagery. We first show a set of results and insights on our initial application of Deep VPR to shore-line imagery and compare these to traditional land-based locational imagery. Using a novel image salience technique to highlight what specific key features in each of the two categories our CNN architecture is picking up on. As this work was carried out during COVID our in-house dataset could not be generated at the time and as such the Symphony Lake dataset whose images are somewhat shoreline-adjacent is used as a stand-in. Secondly, with our in-house dataset then available, we carry out a series of experiments based around the modification of a state-of-the-art Deep VPR pipeline in order to exploit salient feature regions in Shoreline Imagery, as well as tackle the issue of feature redundancy. Our experiments lead us to a novel domain-specific pipeline that provides new state-of-the-art results on our in-house dataset. This pipeline is then used as the basis for a novel human-centered study analysing trust in Deep VPR for Shoreline imagery; the study is made up of several independent surveys including a control survey that simply shows a series of image matching results from the pipeline, a second survey making use of our previously discussed saliency visualisations to communicate model feature extraction to the user explicitly, and surveys that allow the user to intervene in the models decision making directly. The outcome of this work is to show how Deep VPR translates to the shoreline image domain, how the features of this domain differ from land-based imagery, and how we can build pipelines to take advantage of these differences to achieve better results, and how we can ensure user trust in these Deep VPR pipelines under real-life navigation scenarios using various human-computer interaction techniques. |
|---|---|
| Item Description: |
A selection of content is redacted or is partially redacted from this thesis to protect sensitive and personal information. |
| Keywords: |
Place Recognition, Waterborne Imagery, Region Proposal, Image Segmentation, Unsupervised Learning |
| College: |
Faculty of Science and Engineering |
| Funders: |
EPSRC CDT for Enhancing Human Collaborations and Interactions with Data and Intelligence-Driven Systems |