Pose estimation, which means detecting the position and orientation of an object, in space plays a crucial role in many operations, such as docking, debris removal, and inter-spacecraft communications. With the rapid development of neural network solutions in recent years, Deep Learning has been shown to outperform classical computer vision methods in various tasks, including image segmentation or object detection. Therefore, Deep Learning techniques brought significant advances and performance gains in pose estimation tasks. This thesis is inserted in this landscape and its aim is to develop a model-based technique built on Deep Learning to classify an object and determine its pose by comparing it to simple three dimensional geometries. Firstly, a brief overview on Machine Learning is presented, then the frameworks of the implemented Neural Networks are illustrated. In second instance, this thesis shows the process of how synthetic images were generated. Then the results of the training and testing of the neural networks are described. Eventually, the limitations of the model and some possible future developments are reported.
Pose Estimation è un termine che indica rilevare la posizione e l'orientazione di un oggetto e nello spazio gioca un ruolo cruciale in molte operazioni, come il docking, la rimozione dei detriti o le comunicazioni tra spacecrafts. Con il rapido e recente sviluppo delle reti neurali, il Deep Learning si è dimostrato più performante dei metodi classici di computer vision in vari ambiti, tra i quali la segmentazione di un'immagine e l'individuazione di un oggetto. Di conseguenza il Deep Learning ha portato vantaggi significativi anche nel settore della pose estimation. Il lavoro presentato in questo tesi si inserisce in questo panorama e il suo scopo è sviluppare una tecnica di Deep Learning basata su dei modelli per classificare un oggetto e determinarne la posa comparandola con solidi geometricamente più semplici. Innanzitutto la tesi illustra velocemente cosa sia il Machine Learning e poi presenta le archittetture delle reti neurali implementate. In seguito viene illustrato il processo di generazione delle immagini sintetiche. Dopo ciò, vengono riportati i risultati dell'allenamento e della validazione delle reti neurali. Infine vengono illustrate le limitazioni del modello e possibili sviluppi futuri
Convolutional neural network for model based pose estimation
VERDELLI, ENRICO
2020/2021
Abstract
Pose estimation, which means detecting the position and orientation of an object, in space plays a crucial role in many operations, such as docking, debris removal, and inter-spacecraft communications. With the rapid development of neural network solutions in recent years, Deep Learning has been shown to outperform classical computer vision methods in various tasks, including image segmentation or object detection. Therefore, Deep Learning techniques brought significant advances and performance gains in pose estimation tasks. This thesis is inserted in this landscape and its aim is to develop a model-based technique built on Deep Learning to classify an object and determine its pose by comparing it to simple three dimensional geometries. Firstly, a brief overview on Machine Learning is presented, then the frameworks of the implemented Neural Networks are illustrated. In second instance, this thesis shows the process of how synthetic images were generated. Then the results of the training and testing of the neural networks are described. Eventually, the limitations of the model and some possible future developments are reported.File | Dimensione | Formato | |
---|---|---|---|
Tesi Verdelli.pdf
solo utenti autorizzati dal 23/11/2022
Dimensione
10.33 MB
Formato
Adobe PDF
|
10.33 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/10589/182026