The registration of 3D point clouds is a central problem in robotics, computer vision, and 3D graphics, made even more relevant by the growing adoption of LiDAR sensors and BIM models. In this context, the objective of this work is to develop a lightweight self-supervised pipeline, named SPIRALE (Self-supervised Point-cloud Iterative Registration with RGB-Aware Lightweight Auto-Encoder), capable of registering point clouds by simultaneously minimizing both geometric and color errors. SPIRALE combines spatial point sampling and latent representation extraction through a lightweight Auto-Encoder with a cost function that is sensitive to color information. The Auto-Encoder provides compact and discriminative descriptions that support robust and accurate global iterative alignment, even in the presence of noise and low overlap. Experimental results show that SPIRALE achieves significantly higher accuracy compared to both traditional geometric methods and generic deep learning approaches, while maintaining reduced execution times and low computational requirements. Some residual sensitivity to rotations is still observed. In conclusion, SPIRALE represents a convincing balance between accuracy and efficiency, with potential applications in scenarios such as data labeling, mobile industrial inspection, and localization for autonomous vehicles. This approach lays the groundwork for future developments in the field of autonomous and efficient 3D registration.
La registrazione di nuvole di punti 3D è un problema centrale in robotica, visione artificiale e grafica 3D, reso ancora più rilevante dall’adozione crescente di sensori LiDAR e modelli BIM. In questo contesto, l’obiettivo di questo lavoro è sviluppare una pipeline auto-supervisionata leggera, denominata SPIRALE (Self-supervised Point-cloud Iterative Registration with RGB-Aware Lightweight Auto-Encoder), capace di registrare nuvole di punti minimizzando contemporaneamente l’errore geometrico e quello cromatico. SPIRALE combina campionamento spaziale dei punti ed estrazione di rappresentazioni latenti tramite un Auto-Encoder leggero e una funzione di costo sensibile alle informazioni cromatiche. L’Auto-Encoder, fornisce descrizioni compatte e discriminative che supportano un allineamento globale iterativo robusto e accurato, anche in presenza di rumore e bassa sovrapposizione. I risultati sperimentali dimostrano che SPIRALE ottiene un’accuratezza significativamente superiore rispetto sia ai metodi geometrici tradizionali sia agli approcci deep learning generici, mantenendo al contempo tempi di esecuzione ridotti e requisiti computazionali contenuti. Si riscontra ancora una sensibilità residua alle rotazioni. In conclusione, SPIRALE rappresenta un convincente equilibrio tra accuratezza ed efficienza, con potenziali impatti applicativi in scenari quali etichettatura dei dati, l’ispezione industriale e la localizzazione per veicoli autonomi. Tale approccio pone le basi per futuri sviluppi nel campo della registrazione 3D autonoma ed efficiente.
SPIRALE: self-supervised point cloud iterative registration with Rgb-aware lightweight deep auto-encoder
Dominici, Riccardo
2024/2025
Abstract
The registration of 3D point clouds is a central problem in robotics, computer vision, and 3D graphics, made even more relevant by the growing adoption of LiDAR sensors and BIM models. In this context, the objective of this work is to develop a lightweight self-supervised pipeline, named SPIRALE (Self-supervised Point-cloud Iterative Registration with RGB-Aware Lightweight Auto-Encoder), capable of registering point clouds by simultaneously minimizing both geometric and color errors. SPIRALE combines spatial point sampling and latent representation extraction through a lightweight Auto-Encoder with a cost function that is sensitive to color information. The Auto-Encoder provides compact and discriminative descriptions that support robust and accurate global iterative alignment, even in the presence of noise and low overlap. Experimental results show that SPIRALE achieves significantly higher accuracy compared to both traditional geometric methods and generic deep learning approaches, while maintaining reduced execution times and low computational requirements. Some residual sensitivity to rotations is still observed. In conclusion, SPIRALE represents a convincing balance between accuracy and efficiency, with potential applications in scenarios such as data labeling, mobile industrial inspection, and localization for autonomous vehicles. This approach lays the groundwork for future developments in the field of autonomous and efficient 3D registration.| File | Dimensione | Formato | |
|---|---|---|---|
|
Executive Summary Dominici.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive Summary
Dimensione
1.15 MB
Formato
Adobe PDF
|
1.15 MB | Adobe PDF | Visualizza/Apri |
|
Dominici Thesis.pdf
non accessibile
Descrizione: Tesi di Laurea Magistrale
Dimensione
7.31 MB
Formato
Adobe PDF
|
7.31 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/10589/240711