Simultaneous Localization And Mapping (SLAM) is a system that allows a robot, or generic mobile device, to construct a map of its surroundings and at the same time define its pose (position and rotation), without a priori information. In particular, when this process is performed solely through the use of cameras, it is called Visual-SLAM. In this thesis, the Visual-SLAM ORB-SLAM2 system is implemented with the ultimate goal of demonstrating the validity of employing virtual simulations as a testing environment. Two constituent blocks can be identified in the paper. The first part focuses first on the introduction of the SLAM problem, from its probabilistic formulation to the general characteristics of a method based on vision systems, and then moves on to a detailed description of the used implementation of the ORBSLAM2 system. The intention is to provide a complete and in-depth map of the algorithm: from input image processing to loop closure recognition and subsequent optimization. The second part shows how the framework was prepared to carry out the simulations and what results were obtained. A series of tests demonstrate the usefulness of this tool for quantifying the effect of some typical conditions to which a Visual-SLAM system is subjected, such as the presence of dynamic objects, adverse lighting conditions, and high speed of the camera motion. A comparison is then presented between the outcome of the test performed using the KITTI public dataset and the outcome of the tests done in the virtual environments.
La localizzazione e mappatura simultanea (Simultaneous Localization And Mapping, SLAM) è un sistema che permette ad un robot, o generico dispositivo mobile, di costruire una mappa dell’ambiente circostante e allo stesso tempo di definire la propria posa (posizione e rotazione), senza informazioni a priori. In particolare, quando questo processo viene realizzato solamente mediante l’uso di telecamere, esso prende il nome di Visual-SLAM. In questa tesi viene implementato il sistema Visual-SLAM ORB-SLAM2 con lo scopo ultimo di dimostrare la validità di impiego di simulazioni virtuali come ambiente di testing. Possono essere individuati due blocchi costituitivi nell’elaborato. La prima parte si concentra prima sull’introdurre il problema SLAM, dalla sua formulazione probabilistica alle caratteristiche generali di un metodo basato su sistemi di visione, per poi spostarsi ad una descrizione dettagliata della implementazione utilizzata del sistema ORB-SLAM2. L’intenzione è di fornire una completa e approfondita mappa dell’algoritmo: dalla elaborazione delle immagini in input fino al riconoscimento della chiusura del loop e successiva ottimizzazione. La seconda parte illustra come è stato preparato il framework per effettuare le simulazioni e quali sono i risultati ottenuti. Una serie di test dimostra l’utilità di questo strumento per quantificare l’effetto di alcune condizioni tipiche a cui è soggetto un sistema Visual-SLAM, quali la presenza di oggetti dinamici, condizioni di illuminazione ed elevata velocità del moto delle telecamere. Viene poi presentato un confronto tra l’esito del test sostenuto utilizzando il dataset pubblico KITTI e l’esito dei test fatti negli ambienti virtuali.
Analysis of a stereo visual SLAM system in a virtual and real environment
SABBADIN, DAVIDE
2022/2023
Abstract
Simultaneous Localization And Mapping (SLAM) is a system that allows a robot, or generic mobile device, to construct a map of its surroundings and at the same time define its pose (position and rotation), without a priori information. In particular, when this process is performed solely through the use of cameras, it is called Visual-SLAM. In this thesis, the Visual-SLAM ORB-SLAM2 system is implemented with the ultimate goal of demonstrating the validity of employing virtual simulations as a testing environment. Two constituent blocks can be identified in the paper. The first part focuses first on the introduction of the SLAM problem, from its probabilistic formulation to the general characteristics of a method based on vision systems, and then moves on to a detailed description of the used implementation of the ORBSLAM2 system. The intention is to provide a complete and in-depth map of the algorithm: from input image processing to loop closure recognition and subsequent optimization. The second part shows how the framework was prepared to carry out the simulations and what results were obtained. A series of tests demonstrate the usefulness of this tool for quantifying the effect of some typical conditions to which a Visual-SLAM system is subjected, such as the presence of dynamic objects, adverse lighting conditions, and high speed of the camera motion. A comparison is then presented between the outcome of the test performed using the KITTI public dataset and the outcome of the tests done in the virtual environments.File | Dimensione | Formato | |
---|---|---|---|
Sabbadin_Davide.pdf
accesso aperto
Dimensione
18.06 MB
Formato
Adobe PDF
|
18.06 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/43389