Enhancing autonomous vision-based navigation in space with deep learning technologies
Enhancing autonomous vision-based navigation in space with deep learning technologies
Due to the steadily increasing quantity of debris in orbit, active debris removal is widely considered to be a necessary technology; however, no such missions have yet been flown. This is largely down to the challenges faced by the guidance, navigation and control system for close-range rendezvous, and the risks involved. In this thesis, deep learning technologies are applied to the problem of autonomous visual guidance in space, with the aim of increasing the technology readiness level of active debris removal. The key difficulty involved with applying deep learning technologies to this problem is the lack of freely available training data. Therefore, a simulation framework is constructed to generate labelled image datasets for training deep learning models. Both the datasets and software are publicly released in order to facilitate further research in this domain. An approach for determining the attitude evolution of an unknown and uncooperative debris object is proposed, based upon a siamese convolutional neural network. The network detects and tracks inherently useful features from visual sensor data, from which the rotational state is inferred. The approach is fully end-to-end trainable, including an adapted outlier rejection algorithm, allowing for simple finetuning and application to different situations. The method is analysed numerically using simulated data, displaying improved performance compared with state-of-the-art algorithmic approaches. It is also shown to be capable of real-time performance while generalising well to unseen targets. This approach can therefore be considered a feasible solution for safely performing guidance and navigation in active debris removal, satellite servicing and other close proximity operations around a previously unknown object in space. Finally, this research is placed within the context of developing a fully autonomous, artificial intelligence-based on-board guidance and navigation system. Several potential further applications of deep learning to the problem are discussed, along with the remaining challenges faced by these technologies.
University of Southampton
Guthrie, Ben
03652589-afe9-4d34-a4fd-831cb1562c73
August 2022
Guthrie, Ben
03652589-afe9-4d34-a4fd-831cb1562c73
Kim, Min Kwan
18ed9a6f-484f-4a7c-bf24-b630938c1acc
Guthrie, Ben
(2022)
Enhancing autonomous vision-based navigation in space with deep learning technologies.
University of Southampton, Doctoral Thesis, 174pp.
Record type:
Thesis
(Doctoral)
Abstract
Due to the steadily increasing quantity of debris in orbit, active debris removal is widely considered to be a necessary technology; however, no such missions have yet been flown. This is largely down to the challenges faced by the guidance, navigation and control system for close-range rendezvous, and the risks involved. In this thesis, deep learning technologies are applied to the problem of autonomous visual guidance in space, with the aim of increasing the technology readiness level of active debris removal. The key difficulty involved with applying deep learning technologies to this problem is the lack of freely available training data. Therefore, a simulation framework is constructed to generate labelled image datasets for training deep learning models. Both the datasets and software are publicly released in order to facilitate further research in this domain. An approach for determining the attitude evolution of an unknown and uncooperative debris object is proposed, based upon a siamese convolutional neural network. The network detects and tracks inherently useful features from visual sensor data, from which the rotational state is inferred. The approach is fully end-to-end trainable, including an adapted outlier rejection algorithm, allowing for simple finetuning and application to different situations. The method is analysed numerically using simulated data, displaying improved performance compared with state-of-the-art algorithmic approaches. It is also shown to be capable of real-time performance while generalising well to unseen targets. This approach can therefore be considered a feasible solution for safely performing guidance and navigation in active debris removal, satellite servicing and other close proximity operations around a previously unknown object in space. Finally, this research is placed within the context of developing a fully autonomous, artificial intelligence-based on-board guidance and navigation system. Several potential further applications of deep learning to the problem are discussed, along with the remaining challenges faced by these technologies.
Text
PhD_Thesis_Final_Submission
- Version of Record
Text
PTD_Thesis_Guthrie-SIGNED
Restricted to Repository staff only
More information
Published date: August 2022
Identifiers
Local EPrints ID: 470293
URI: http://eprints.soton.ac.uk/id/eprint/470293
PURE UUID: d6d333fd-ff99-466e-bcb3-11d4b5c38979
Catalogue record
Date deposited: 05 Oct 2022 16:54
Last modified: 17 Mar 2024 03:33
Export record
Contributors
Author:
Ben Guthrie
Download statistics
Downloads from ePrints over the past year. Other digital versions may also be available to download e.g. from the publisher's website.
View more statistics