The research on vision-based pose estimation algorithms using monocular images for Rendezvous and Proximity Operations has been widely discussed in recent years. Deep Learning (DL) techniques perform well in such research, but they require labelled spacecraft datasets. The existing datasets contain a single spacecraft, and the algorithms are inaccessible to reproduce the dataset with any desired spacecraft. This research is proposed to overcome these shortcomings by introducing an algorithm to generate a multi-spacecraft dataset named POSSE Dataset. Since the state-of-the-art (SOTA) models are trained to detect the very same or similar featured spacecraft, this research also aims to substantiate these DL models' feasibility using a convolution neural network (CNN) in detecting the spacecraft with disparate features eventually assisting in pose estimation. The dataset generation algorithm written in Python makes use of Relative Orbital Elements for the camera path planning and is developed to integrate with open-source software using physical-based rendering techniques. Moreover, the proposed CNN for detection termed as Spacecraft Detection Network (SDN) employs modified SOTA object detection architectures to predict multiple spacecraft. As a result, the POSSE Dataset includes 8237 novel synthetic images of Aqua and Cluster that consist of sharp and smooth features along with their annotations for a bounding box, keypoints and pose data. In addition, the dataset includes challenging conditions such as long-range, low visibility, strong shadows, and partial views of the spacecraft with stars and Earth in the background. Furthermore, the POSSE Dataset is partitioned into another four datasets by removing the stars in the background to validate its robustness in comparison, with the proposed SDN in different scenarios. Overall, the SDN is proven to be efficient in detecting a single spacecraft with a dark background at a mean Average Precision (mAP) of 78.03% and 5 Frames Per Second (FPS). However, this performance is reduced when tested on the POSSE Dataset to 36.57% of mAP and 4 FPS.
La ricerca sugli algoritmi di stima della posa basati sulla visione utilizzando immagini monoculari per le Operazioni di Rendezvous e Prossimità è stata ampiamente discussa negli ultimi anni. Le tecniche di Deep Learning (DL) danno buoni risultati, ma richiedono dataset etichettati dei satelliti. I dataset esistenti contengono un solo satellite, e gli algoritmi sono inaccessibili per riprodurre i dataset con qualsiasi satellite desiderato. Lo scopo di questa ricerca è superare questi svantaggi introducendo un algoritmo per generare un dataset multi-satellite chiamato POSSE Dataset. Poiché i modelli allo stato dell'arte (SOTA) sono addestrati per rilevare lo stesso satellite o simili, questa ricerca mira anche a corroborare la loro fattibilità di DL utilizzando una rete neurale convoluzionale (CNN) nella rilevazione dei satelliti con caratteristiche disparate, aiutando infine nella stima della posa. L'algoritmo di generazione del dataset scritto in Python fa uso degli Elementi Orbitali Relativi per la pianificazione del percorso della telecamera ed è sviluppato per integrarsi con software open-source utilizzando tecniche di rendering basate sulla fisica. Inoltre, la CNN proposta per la rilevazione, denominata Spacecraft Detection Network (SDN), impiega architetture di rilevamento oggetti modificate dallo SOTA per prevedere più satelliti. Di conseguenza, il POSSE Dataset include 8237 nuove immagini sintetiche di Aqua e Cluster che consistono in caratteristiche nitide e lisce insieme alle loro annotazioni per bounding box, keypoints e dati di posa. Inoltre, il dataset include condizioni difficili come lunghe distanze, bassa visibilità, forti ombre e viste parziali dei satelliti, con stelle e la Terra sullo sfondo. Inoltre, il POSSE Dataset è suddiviso in altri quattro dataset rimuovendo le stelle sullo sfondo per convalidare la sua robustezza in confronto con la proposta SDN in scenari diversi. Complessivamente, la SDN è dimostrata efficiente nel rilevare un singolo satellite su sfondo scuro con una mean Average Precision (mAP) del 78,03% e 5 Frame Per Second (FPS). Tuttavia, questa prestazione è ridotta quando testata sul POSSE Dataset al 36,57% di mAP e 4 FPS.
Multi-class Dataset Generation and Spacecraft Detection with Deep Learning to aid Pose Estimation
Bhupanam, Eshwar Bhargav
2022/2023
Abstract
The research on vision-based pose estimation algorithms using monocular images for Rendezvous and Proximity Operations has been widely discussed in recent years. Deep Learning (DL) techniques perform well in such research, but they require labelled spacecraft datasets. The existing datasets contain a single spacecraft, and the algorithms are inaccessible to reproduce the dataset with any desired spacecraft. This research is proposed to overcome these shortcomings by introducing an algorithm to generate a multi-spacecraft dataset named POSSE Dataset. Since the state-of-the-art (SOTA) models are trained to detect the very same or similar featured spacecraft, this research also aims to substantiate these DL models' feasibility using a convolution neural network (CNN) in detecting the spacecraft with disparate features eventually assisting in pose estimation. The dataset generation algorithm written in Python makes use of Relative Orbital Elements for the camera path planning and is developed to integrate with open-source software using physical-based rendering techniques. Moreover, the proposed CNN for detection termed as Spacecraft Detection Network (SDN) employs modified SOTA object detection architectures to predict multiple spacecraft. As a result, the POSSE Dataset includes 8237 novel synthetic images of Aqua and Cluster that consist of sharp and smooth features along with their annotations for a bounding box, keypoints and pose data. In addition, the dataset includes challenging conditions such as long-range, low visibility, strong shadows, and partial views of the spacecraft with stars and Earth in the background. Furthermore, the POSSE Dataset is partitioned into another four datasets by removing the stars in the background to validate its robustness in comparison, with the proposed SDN in different scenarios. Overall, the SDN is proven to be efficient in detecting a single spacecraft with a dark background at a mean Average Precision (mAP) of 78.03% and 5 Frames Per Second (FPS). However, this performance is reduced when tested on the POSSE Dataset to 36.57% of mAP and 4 FPS.| File | Dimensione | Formato | |
|---|---|---|---|
|
2023_12_Bhupanam.pdf
accessibile in internet per tutti
Descrizione: Thesis report
Dimensione
36.42 MB
Formato
Adobe PDF
|
36.42 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/215630