Full Text Available
Note: Clicking the button above will open the full text document at the original institutional repository in a new window.
Pose estimation of space objects with neural networks
Thesis (MEng)--Stellenbosch University, 2024.
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | en_ZA en_ZA |
| Published: |
Stellenbosch : Stellenbosch University
2024
|
| Subjects: | |
| Tags: |
No Tags, Be the first to tag this record!
|
| _version_ | 1867614096715677696 |
|---|---|
| access_status_str | Open Access |
| author | Korf, Gerhardus |
| author2 | Jordaan, Willem |
| author_browse | Jordaan, Willem Korf, Gerhardus |
| author_facet | Jordaan, Willem Korf, Gerhardus |
| author_sort | Korf, Gerhardus |
| collection | Thesis |
| dc_rights_str_mv | Stellenbosch University |
| description | Thesis (MEng)--Stellenbosch University, 2024. |
| format | Thesis |
| id | oai:scholar.sun.ac.za:10019.1/130329 |
| institution | Stellenbosch University (South Africa) |
| language | en_ZA en_ZA |
| last_indexed | 2026-06-10T12:46:36.532Z |
| license_str | Other — see source repository |
| provenance_str_mv | Harvested via OAI-PMH from SUNScholar — Stellenbosch University Repository |
| publishDate | 2024 |
| publishDateRange | 2024 |
| publishDateSort | 2024 |
| publisher | Stellenbosch : Stellenbosch University |
| publisherStr | Stellenbosch : Stellenbosch University |
| record_format | dspace |
| source_str | SUNScholar — Stellenbosch University Repository |
| spelling | oai:scholar.sun.ac.za:10019.1/130329 Pose estimation of space objects with neural networks Korf, Gerhardus Jordaan, Willem Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering. Space debris Demodulation (Electronics) Computer vision Neural networks (Computer science) UCTD Thesis (MEng)--Stellenbosch University, 2024. ENGLISH ABSTRACT: The recent rise in the commercial use of space has led to a surge in space debris such as defunct satellites and rocket bodies. This poses a threat to future space expeditions as more space debris escalates the risk of unwanted collisions within orbit. Active Debris Removal (ADR) missions aim to reduce the amount of debris by launching spacecrafts that rendezvous with targets and deploy deorbiting plans for them. ADR missions are inherently complex with technical challenges that still need to be overcome. One of these challenges is that an ADR spacecraft requires knowledge of the target object’s pose so that it can interact with it under the lowest risk circumstances. This knowledge often has to be estimated as debris can be uncooperative and unable to communicate with the ADR spacecraft. This project investigates the viability of using Neural Networks (NN) to estimate the pose of targets within orbit. Cube Satellites (CubeSats) are used as the debris targets and monocular cameras are used to capture them. A robust NN-based solution that estimates pose from just the images captured of the CubeSat and assumed knowledge about it is developed. The literature on pose estimation within orbit shows that large and asymmetrical targets are often used, which makes the research of NN-based pose estimation solutions for a small and typically symmetrical target such as a CubeSat valuable. Furthermore, traditional control theory methods are often omitted from solutions that rely on machine learning techniques. Thus, the investigation of the inclusion of an Extended Kalman Filter (EKF) within solutions is valuable as well. NNs require data to be developed. This is a problem as there is a lack of real-life data of specific CubeSats being captured within orbit. As a practical replacement, synthetic data of CubeSats are created with the Three-Dimensional (3D) rendering software Unreal Engine (UE). The data is made as realistic as possible with the intended functionality of using it to train NNs, and then applying the NNs to real-life data. From qualitative analyses, it is clear that the created synthetic data is visually similar to reality. A range of potential NN-based solutions are designed and implemented. The solutions range from being end-to-end, where a NN is solely relied upon to predict the pose of a CubeSat, or partial, where a NN is used as part of a greater solution that predicts the pose of a CubeSat. Alongside these solutions, an EKF that models an ADR spacecraft and CubeSat encounter within orbit is created. The EKF can be used to refine pose predictions from the potential solutions, or to estimate the pose itself. The solutions, along with the EKF, are evaluated on a simulated CubeSat. A partial solution that relies on a NN detecting an Attitude Independent Bounding Box (AIBB) of the CubeSat is determined to be the best solution for predicting translation. The inclusion of an EKF within this solution reduces its error significantly and indicates its value. Another partial solution that uses the Two-Dimensional (2D) coordinates of detected landmarks on the surface of a CubeSat as well as a corresponding set of 3D coordinates as measurements in an EKF, was determined to be the best solution for estimating attitude. The two partial solutions are combined into a single robust pose estimation solution. To address the use of synthetically trained NNs in real-life, the robust pose estimation solution is applied to an imitation CubeSat. The CubeSat is constructed in real-life and recreated in UE in order to create synthetic data and train the NNs within the solution. Real-life images of the CubeSat are taken to evaluate the NNs and the solution. The performance of the solution is somewhat weakened when comparing it to the performance achieved on the simulated target. This is determined to be caused by the simulation to reality jump and synthetic data that does not recreate real-life perfectly. Despite a drop in performance, the solution still performs reasonably well under its intended circumstances. By increasing the realism of the synthetic data used to train the NNs, it might be possible to lessen this drop in performance. From the evaluations on the simulated and real-life CubeSats, it is shown that NNs are a valuable tool in the context of pose estimation. The inclusion of an EKF is also valuable as it allows the solution to be robust to noise and erroneous intermediate predictions by the NNs. AFRIKAANSE OPSOMMING: Die onlangse toename in die kommersi¨ele gebruik van die ruimte het gelei tot ’n styging in ruimterommel soos stukkende satelliete en dele van ou vuurpyle. Hierdie is ’n bedreiging vir toekomstige ruimte-ekspedisies aangesien meer ruimterommel die risiko van ongewenste botsings in die ruimte verhoog. Aktiewe Ruimterommelverwydering (ADR)-missies se doel is om die hoeveelheid rommel te verlaag deur ruimtetuie te lanseer wat teikens in die ruimte help om uit hulle wentelbane te ontsnap en terug na die Aarde te val. ADRmissies is ingewikkeld en het tegniese uitdagings wat nog opgelos moet word. Een van hierdie uitdagings is dat ’n ADR-ruimtetuig kennis van ’n teikenvoorwerp se posisie en ori¨entasie benodig sodat dit onder die laagste risikoomstandighede aksies daarop kan uitvoer. Hierdie kennis moet dikwels geskat word omdat rommel onsamewerkend kan wees en nie noodwendig met die ADR-ruimtetuig kan kommunikeer nie. Hierdie projek ondersoek die lewensvatbaardheid van die gebruik van Neurale Netwerke (NN’s) om posisie en ori¨entasie te skat in die ruimte. CubeSats word as rommelteikens gebruik en eenoogkameras word gebruik om hulle af te neem. ’n Robuuste NN-gebaseerde oplossing wat posisie en ori¨entasie skat van slegs beelde van die CubeSat and aangenome kennis daarvan, word ontwikkel. Die literatuur oor hierdie probleem toon aan dat groot en asimmetriese teikens dikwels gebruik word, wat navorsing oor NN-gebaseerde oplossings vir ’n klein en dikwels simmetriese teiken soos ’n CubeSat waardevol maak. Verder word tradisionele beheerteorie-metodes dikwels weggelaat uit oplossings wat op masjienleertegnieke staatmaak. Daarom is die ondersoek van die insluiting van ’n Uitgebreide Kalman Filter (EKF) binne oplossings ook waardevol. NN’s vereis data om ontwikkel te word. Hierdie is ’n probleem aangesien daar ’n gebrek is aan werklike data van spesifieke CubeSats wat in die ruimte afgeneem word. As ’n praktiese vervanging word sintetiese data van CubeSats geskep met die Driedimensionele (3D) visuele sagteware Unreal Engine (UE). Die data word so realisties as moontlik gemaak met die gewenste funksionalitiet van dit gebruik om NN’s te leer, en dan die NN’s op werklike data toe te pas. Van kwalitatiewe ontledings blyk dit dat die geskepte sintetiese data visueel soortgelyk aan die werklikheid is. ’n Verskeidenheid van moontlike NN-gebaseerde oplossings word ontwerp en ge¨ımplementeer. Die oplossings wissel van einde-tot-einde, waar ’n NN uitsluitlik gebruik word om die posisie en ori¨entasie van ’n CubeSat te voorspel, tot gedeeltelik, waar ’n NN deel is van ’n groter oplossing wat die posisie en ori¨entasie van ’n CubeSat voorspel. ’n EKF wat ’n ADR-ruimtetuig en CubeSat ontmoeting in die ruimte modelleer, word ook geskep. Die EKF kan gebruik word om posisie en ori¨entasie voorspellings te verwyn, of self die posisie en ori¨entasie te skat. Die oplossings, tesame met die EKF, word ge¨evalueer op ’n gesimuleerde CubeSat. ’n Gedeeltelike oplossing wat staatmaak op ’n NN wat ’n Ori¨entasie Onafhanklike Grenskas (AIBB) identifiseer, word as die beste oplossing vir posisievoorspelling gevind. Die insluiting van ’n EKF in hierdie oplossing verminder sy fout aansienlik, wat dui daarop dat dit waardevol is. ’n Ander gedeeltelike oplossing wat die Tweedimensionele (2D) ko¨ordinate van ge¨ıdentifiseerde landmerke op die oppervlak van ’n CubeSat, sowel as ’n ooreenstemmende stel van 3D ko¨ordinate as metings in ’n EKF gebruik, is as die beste oplossing vir ori¨entasieskatting gevind. Die twee gedeeltelike oplossings word saamgesit in ’n enkele robuuste oplossing wat beide posisie en ori¨entasie skat. Om die gebruik van sintetiese geleerde NN’s in die werklike lewe aan te spreek, word die robuuste oplossing toegepas op ’n nagebootse CubeSat. Die CubeSat word in werklikheid gebou en in UE herskep om sintetiese data te maak en die NN’s in die oplossing te leer. Werklike beelde van die CubeSat word afgneem om die NN’s en die oplossing te evalueer. Die fout van die oplossing is effens meer as dit wat gesien is met die gesimuleerde teiken. Dit word bepaal dat hierdie veroorsaak word deur die oorgang van simulasie na werklikheid en sintetiese data wat nie die werklikheid perfek herskep nie. Ten spyte van hierdie verhoging in die fout, werk die oplossing steeds redelik goed onder sy beoogde omstandighede. Deur die realisme van die sintetiese data wat gebruik word om die NN’s te leer te verhoog, mag dit moontlik wees om hierdie toename in fout te verminder. Uit die evaluasies op die gesimuleerde en werklike CubeSats blyk dit dat NN’s ’n waardevolle hulpmiddel is vir die skatting van posisie en ori¨entasie. Die insluting van ’n EKF is ook waardevol omdat dit die oplossing in staat stel om robuust te wees teen ruis en foutiewe tussenliggende voorspellings deur die NN’s. Masters 2024-02-29T07:23:47Z 2024-04-26T13:38:50Z 2024-02-29T07:23:47Z 2024-04-26T13:38:50Z 2024-03 Thesis https://scholar.sun.ac.za/handle/10019.1/130329 en_ZA en_ZA Stellenbosch University xvii, 124 pages : illustrations. application/pdf Stellenbosch : Stellenbosch University |
| spellingShingle | Space debris Demodulation (Electronics) Computer vision Neural networks (Computer science) UCTD Korf, Gerhardus Pose estimation of space objects with neural networks |
| title | Pose estimation of space objects with neural networks |
| title_full | Pose estimation of space objects with neural networks |
| title_fullStr | Pose estimation of space objects with neural networks |
| title_full_unstemmed | Pose estimation of space objects with neural networks |
| title_short | Pose estimation of space objects with neural networks |
| title_sort | pose estimation of space objects with neural networks |
| topic | Space debris Demodulation (Electronics) Computer vision Neural networks (Computer science) UCTD |
| url | https://scholar.sun.ac.za/handle/10019.1/130329 |
| work_keys_str_mv | AT korfgerhardus poseestimationofspaceobjectswithneuralnetworks |