Event cameras are a novel class of vision sensors that operate fundamentally differently from traditional frame-based cameras. Instead of capturing images at fixed intervals, they asynchronously record changes in pixel intensity as individual events, triggered when a predefined brightness threshold is exceeded. This unique data acquisition paradigm enables extremely high temporal resolution, low latency, low power consumption, and an inherently high dynamic range, making event cameras particularly well-suited for use in visually and dynamically challenging environments. One of their primary applications is in Simultaneous Localization and Mapping (SLAM), where the goal is to estimate the trajectory of a moving camera while simultaneously reconstructing a 3D map of the environment. In this thesis, Gaussian Splatting, which was originally developed for photorealistic 3D scene reconstruction, is integrated into an event-based SLAM pipeline with a particular focus on space scenarios such as autonomous docking. By leveraging temporally aggregated event representations and modeling the scene as a collection of volumetric Gaussian primitives, the approach aims to provide robust, accurate, and continuous 3D mapping from sparse visual input, while the system is designed for incremental operation so that it ensures consistent scene reconstruction and reliable pose estimation even under the extreme lighting and motion conditions typical of orbital proximity operations.
Le camere ad eventi rappresentano una nuova generazione di sensori visivi che si distinguono nettamente dalle tradizionali camere a fotogrammi. Invece di catturare immagini a intervalli regolari, registrano in maniera asincrona solo le variazioni di luminosità dei pixel, generando eventi ogni volta che viene superata una certa soglia. Questo meccanismo consente una risoluzione temporale molto elevata, tempi di risposta ridottissimi, basso consumo energetico e un’ampia gamma dinamica, caratteristiche che le rendono particolarmente adatte ad ambienti complessi dal punto di vista visivo e dinamico. Tra i principali campi di applicazione vi è lo SLAM (Simultaneous Localization and Mapping), ovvero la capacità di stimare il percorso di una telecamera in movimento costruendo al contempo una mappa tridimensionale dell’ambiente. In questa tesi viene esplorata l’integrazione del Gaussian Splatting, una tecnica di rendering sviluppata per la ricostruzione fotorealistica di scene 3D, all’interno di una pipeline SLAM basata su eventi, con particolare attenzione a scenari spaziali come le manovre di attracco autonomo. L’approccio proposto sfrutta rappresentazioni temporali aggregate degli eventi e descrive la scena come un insieme di primitive gaussiane volumetriche, con l’obiettivo di ottenere una mappatura 3D continua, accurata e robusta anche a partire da dati visivi molto scarsi. Il sistema è stato pensato per funzionare in modo incrementale, garantendo ricostruzioni consistenti della scena e stime affidabili della traiettoria anche nelle condizioni estreme di illuminazione e movimento che caratterizzano le operazioni in orbita.
Event-based pose estimation and scene reconstruction of resident space objects
Fontana, Piercarlo
2024/2025
Abstract
Event cameras are a novel class of vision sensors that operate fundamentally differently from traditional frame-based cameras. Instead of capturing images at fixed intervals, they asynchronously record changes in pixel intensity as individual events, triggered when a predefined brightness threshold is exceeded. This unique data acquisition paradigm enables extremely high temporal resolution, low latency, low power consumption, and an inherently high dynamic range, making event cameras particularly well-suited for use in visually and dynamically challenging environments. One of their primary applications is in Simultaneous Localization and Mapping (SLAM), where the goal is to estimate the trajectory of a moving camera while simultaneously reconstructing a 3D map of the environment. In this thesis, Gaussian Splatting, which was originally developed for photorealistic 3D scene reconstruction, is integrated into an event-based SLAM pipeline with a particular focus on space scenarios such as autonomous docking. By leveraging temporally aggregated event representations and modeling the scene as a collection of volumetric Gaussian primitives, the approach aims to provide robust, accurate, and continuous 3D mapping from sparse visual input, while the system is designed for incremental operation so that it ensures consistent scene reconstruction and reliable pose estimation even under the extreme lighting and motion conditions typical of orbital proximity operations.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_10_Fontana_Tesi_01.pdf
accessibile in internet per tutti
Descrizione: Testo Tesi
Dimensione
34.89 MB
Formato
Adobe PDF
|
34.89 MB | Adobe PDF | Visualizza/Apri |
|
2025_10_Fontana_ExecutiveSummary_02.pdf
accessibile in internet per tutti
Descrizione: Testo Executive Summary
Dimensione
2.44 MB
Formato
Adobe PDF
|
2.44 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/10589/243048