Augmented Reality (AR) technologies have gained widespread popularity in recent years, leading to the development of devices equipped with cameras and visualization components that can project digital objects into the real world. To achieve this seamlessly, a fundamental requirement is the construction of a 3D model of the surroundings, along with an accurate estimation of the camera’s location within that model. This complex challenge is commonly addressed in computer vision through Simultaneous Localization And Mapping (SLAM) techniques. In the context of AR devices, which are often compact with limited computing power, the efficient implementation of SLAM becomes crucial. The primary objective of this thesis is to focus on the adaptation and optimization of the ORB-SLAM3 algorithm for deployment on two specific devices, Jetson Agx Orin and Jetson Nano Orin. The study aims to evaluate the trade-offs between computation time and precision, especially considering the constraints imposed by the computational limitations of these devices. To achieve this, we utilize a stereo-inertial configuration, which leverages both stereo camera vision and inertial measurements for a comprehensive and robust approach to SLAM. Two versions of the algorithm were created, namely the original version and a GPU-accelerated variant tailored for specific algorithmic components. The research aims to determine the optimal performance of two versions when deployed on individual devices. Through systematic testing, it seeks to identify the version that works most effectively for each specific device configuration. This study contributes to the growing body of research in the field of SLAM and AR technologies, with a particular focus on the practical implementation of these techniques on limited computing power platforms. It provides valuable insights into the benchmarking and optimization of SLAM algorithms, which can inform future research in this area.
Le tecnologie di Realtà Aumentata (AR) hanno guadagnato una diffusa popolarità negli ultimi anni, portando allo sviluppo di dispositivi dotati di telecamere e componenti di visualizzazione in grado di proiettare oggetti digitali nel mondo reale. Per realizzare ciò in modo fluido, un requisito fondamentale è la costruzione di un modello 3D dell’ambiente circostante, insieme a una stima accurata della posizione della telecamera all’interno di tale modello. Questa sfida complessa è comunemente affrontata attraverso tecniche di Simultaneous Localization And Mapping (SLAM) nella computer vision. Nel contesto dei dispositivi AR, spesso compatti con limitata potenza di calcolo, l’implementazione efficiente di SLAM diventa cruciale. L’obiettivo principale di questa tesi è concentrarsi sull’adattamento e l’ottimizzazione dell’algoritmo ORB-SLAM3 per l’implementazione su due dispositivi specifici, Jetson Agx Orin e Jetson Nano Orin. Lo studio mira a valutare i compromessi tra il tempo di calcolo e la precisione, specialmente considerando i vincoli imposti dalle limitazioni computazionali di tali dispositivi. Per raggiungere questo obiettivo, utilizziamo una configurazione stereo-inerziale, che sfrutta sia la visione della telecamera stereo che le misurazioni inerziali per un approccio completo e robusto a SLAM. Sono state create due versioni dell’algoritmo, ovvero la versione originale e una variante accelerata da GPU adattata per componenti algoritmiche specifiche. La ricerca si concentra sul testare quale delle due versioni funziona meglio per ciascun dispositivo. Attraverso test sistematici, si cerca di identificare la versione che risulta più efficace per ogni configurazione specifica del dispositivo. Questo studio contribuisce al crescente corpo di ricerca nel campo di SLAM e delle tecnologie AR, con un particolare focus sull’implementazione pratica di queste tecniche su piattaforme con limitata potenza di calcolo. Fornisce preziose intuizioni sul benchmarking e l’ottimizzazione degli algoritmi SLAM, che possono informare future ricerche in questo settore.
Evaluating performance and optimization strategies of SLAM on embedded platforms
Piol, Federico
2023/2024
Abstract
Augmented Reality (AR) technologies have gained widespread popularity in recent years, leading to the development of devices equipped with cameras and visualization components that can project digital objects into the real world. To achieve this seamlessly, a fundamental requirement is the construction of a 3D model of the surroundings, along with an accurate estimation of the camera’s location within that model. This complex challenge is commonly addressed in computer vision through Simultaneous Localization And Mapping (SLAM) techniques. In the context of AR devices, which are often compact with limited computing power, the efficient implementation of SLAM becomes crucial. The primary objective of this thesis is to focus on the adaptation and optimization of the ORB-SLAM3 algorithm for deployment on two specific devices, Jetson Agx Orin and Jetson Nano Orin. The study aims to evaluate the trade-offs between computation time and precision, especially considering the constraints imposed by the computational limitations of these devices. To achieve this, we utilize a stereo-inertial configuration, which leverages both stereo camera vision and inertial measurements for a comprehensive and robust approach to SLAM. Two versions of the algorithm were created, namely the original version and a GPU-accelerated variant tailored for specific algorithmic components. The research aims to determine the optimal performance of two versions when deployed on individual devices. Through systematic testing, it seeks to identify the version that works most effectively for each specific device configuration. This study contributes to the growing body of research in the field of SLAM and AR technologies, with a particular focus on the practical implementation of these techniques on limited computing power platforms. It provides valuable insights into the benchmarking and optimization of SLAM algorithms, which can inform future research in this area.File | Dimensione | Formato | |
---|---|---|---|
2024_04_Piol_Thesis.pdf
non accessibile
Descrizione: file tesi
Dimensione
9.3 MB
Formato
Adobe PDF
|
9.3 MB | Adobe PDF | Visualizza/Apri |
2024_04_Piol_Executive Summary.pdf
non accessibile
Descrizione: file executive summary
Dimensione
2.38 MB
Formato
Adobe PDF
|
2.38 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/10589/217636