This thesis investigates pose-graph estimation on the Lie group SE(d)^n from noisy relative measurements defined on a possibly disconnected measurement graph representing the available relative pose observations. The goal is to estimate a set of unknown absolute poses from relative measurements modeled as Hij = Exp(hat(ξij)) T_i*T_j^(−1), where T_i, T_j ∈ SE(d) are the absolute poses and ξij is a zero-mean heteroscedastic Gaussian noise in the Lie algebra. This measurement model preserves group consistency, ensures noise-free exactness, and is invariant to global right actions on the poses. Due to this invariance, maximum likelihood estimation does not admit a unique solution, but yields a globally consistent set of relative transformations over the measurement graph, enabling coherent inference even in the absence of direct measurements between all poses. A detailed derivation of the Fisher Information Matrix (FIM) is presented using rightinvariant perturbations and Lie group geometry. Two equivalent formulations are obtained: a classical sum of per-measurement contributions and a compact factorization separating the topological structure of the measurement graph from pose-dependent kinematics. The FIM is shown to be singular, with a nullspace dimension equal to c*dim(SE(d)) for a graph with c connected components, reflecting the intrinsic gauge freedom. To derive meaningful Cramér–Rao lower bounds, two approaches are developed. The first introduces anchor poses, one per connected component, restricting the problem to a Riemannian submanifold. The second formulates estimation intrinsically on the quotient manifold SE(d)^n/SE(d)^c. In both cases, useful bounds for unbiased estimators are derived. Numerical simulations on sparse directed pose graphs in SE(2) with heteroscedastic noise validate the theoretical results, confirming the predicted rank deficiency and convergence of the estimator to the intrinsic bounds.
Questa tesi studia il problema della stima di pose sul gruppo di Lie SE(d)^n a partire da misure relative rumorose definite su un grafo delle misure, possibilmente disconnesso, che rappresenta le osservazioni relative di pose disponibili. L’obiettivo è stimare un insieme di pose assolute incognite a partire da misure relative modellate come Hij = Exp(hat(ξij)) T_i*T_j^(−1) , dove T_i, T_j ∈ SE(d) sono le pose assolute e ξij è un rumore gaussiano eteroschedastico a media nulla agente nell’algebra di Lie se(d). Questo modello di misura preserva la consistenza di gruppo, garantisce l’esattezza in assenza di rumore ed è invariante rispetto ad azioni globali a destra sulle pose. A causa di tale invarianza, la stima di massima verosimiglianza non ammette una soluzione unica, ma produce un insieme globalmente consistente di trasformazioni relative sul grafo delle misure, consentendo un’inferenza coerente anche in assenza di misure dirette tra tutte le pose. Viene fornita una derivazione dettagliata della Matrice di Informazione di Fisher (FIM) utilizzando perturbazioni destre e la geometria intrinseca dei gruppi di Lie. Si ottengono due formulazioni equivalenti: una classica, come somma dei contributi di ciascuna misura e una fattorizzazione compatta, che separa la struttura topologica del grafo delle misure dalla dipendenza cinematica delle pose. Si dimostra che la FIM è singolare, con un nucleo di dimensione pari a c*dim(SE(d)) per un grafo con c componenti connesse, riflettendo l’ambiguità di gauge intrinseca. Per derivare limiti inferiori di Cramér–Rao significativi, vengono sviluppati due approcci. Il primo introduce pose ancorate, cioè fissate al valore reale, una per ciascuna componente connessa, restringendo il problema di stima a una sottovarietà riemanniana. Il secondo formula il problema in modo intrinseco sulla varietà quoziente SE(d)^n/SE(d)^c. In entrambi i casi vengono derivati limiti inferiori utili per stimatori non distorti. Simulazioni numeriche su un grafo di pose sparso in SE(2) con rumore eteroschedastico validano i risultati teorici, confermando la deficienza di rango prevista e la convergenza dello stimatore efficiente ai limiti ottenuti.
On intrinsic Cramér-Rao Bounds on the Lie Group SE(d)^n with an application to computer vision
BERTINI, MICHELE
2024/2025
Abstract
This thesis investigates pose-graph estimation on the Lie group SE(d)^n from noisy relative measurements defined on a possibly disconnected measurement graph representing the available relative pose observations. The goal is to estimate a set of unknown absolute poses from relative measurements modeled as Hij = Exp(hat(ξij)) T_i*T_j^(−1), where T_i, T_j ∈ SE(d) are the absolute poses and ξij is a zero-mean heteroscedastic Gaussian noise in the Lie algebra. This measurement model preserves group consistency, ensures noise-free exactness, and is invariant to global right actions on the poses. Due to this invariance, maximum likelihood estimation does not admit a unique solution, but yields a globally consistent set of relative transformations over the measurement graph, enabling coherent inference even in the absence of direct measurements between all poses. A detailed derivation of the Fisher Information Matrix (FIM) is presented using rightinvariant perturbations and Lie group geometry. Two equivalent formulations are obtained: a classical sum of per-measurement contributions and a compact factorization separating the topological structure of the measurement graph from pose-dependent kinematics. The FIM is shown to be singular, with a nullspace dimension equal to c*dim(SE(d)) for a graph with c connected components, reflecting the intrinsic gauge freedom. To derive meaningful Cramér–Rao lower bounds, two approaches are developed. The first introduces anchor poses, one per connected component, restricting the problem to a Riemannian submanifold. The second formulates estimation intrinsically on the quotient manifold SE(d)^n/SE(d)^c. In both cases, useful bounds for unbiased estimators are derived. Numerical simulations on sparse directed pose graphs in SE(2) with heteroscedastic noise validate the theoretical results, confirming the predicted rank deficiency and convergence of the estimator to the intrinsic bounds.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Bertini_Tesi.pdf
accessibile in internet per tutti
Descrizione: Elaborato finale completo
Dimensione
1.78 MB
Formato
Adobe PDF
|
1.78 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Bertini_Executive Summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary richiesto per tesi con controrelatore
Dimensione
433.56 kB
Formato
Adobe PDF
|
433.56 kB | 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/10589/251377