Self-localization of acoustic sensor networks is an important prerequisite for spatial audio applications, but remains challenging when the relative geometry of the network is unknown. This thesis addresses the problem of estimating the full 6-degree-of-freedom relative transformation between two synchronized, non-coplanar microphone arrays from a single acoustic source, without prior knowledge of the room geometry. The proposed approach exploits early acoustic reflections through the image source method, representing them as a 3D point cloud of virtual image sources. The task then amounts to matching and aligning the two point clouds estimated from the arrays, which is made difficult by their noisiness, the presence of spurious points, and the local symmetries produced by the room geometry. To this aim, we introduce ARCTIC (Anchor-based Registration using Congruent Triangle Consensus), a global registration algorithm specifically designed for acoustic image source point clouds. ARCTIC exploits the direct source as a natural correspondence, using it as an anchor to estimate the relative transformation. The method is evaluated on both simulated and real data. The results show that ARCTIC outperforms the considered baselines, reaching up to 97.6% of correct alignments on simulated data according to our evaluation criterion. Evaluated on 36 real measurements, ARCTIC correctly estimates all the transformations, with mean orientation and inter-array distance errors of 4.77° and 9.5 cm, respectively. These results demonstrate the potential of exploiting reverberation and domain-specific acoustic information for microphone-array self-localization from a single acoustic source.
La calibrazione geometrica delle reti di sensori acustici rappresenta un prerequisito importante per le applicazioni di audio spaziale, ma rimane una problematica complessa quando la geometria relativa della rete è sconosciuta. Questa tesi affronta il problema della stima della trasformazione relativa a 6 gradi di libertà tra due array di microfoni sincronizzati e non complanari, a partire da un’unica sorgente acustica e senza una conoscenza preliminare della geometria della stanza. L’approccio proposto sfrutta le riflessioni acustiche attraverso il metodo delle sorgenti immagine, rappresentandole come una nuvola di punti 3D di sorgenti immagine virtuali. Il problema viene quindi ricondotto al matching e all’allineamento delle due nuvole di punti stimate dagli array, un’operazione resa complessa dalla loro rumorosità, dalla presenza di punti spuri e dalle simmetrie locali generate dalla geometria della stanza. A questo scopo, introduciamo ARCTIC (Anchor-based Registration using Congruent Triangle Consensus), un algoritmo di calibrazione globale progettato specificamente per nuvole di punti acustiche generate da sorgenti immagine. ARCTIC sfrutta la sorgente diretta come riferimento naturale, utilizzandola come ancoraggio per stimare la trasformazione relativa. Il metodo viene valutato sia su dati simulati sia su dati reali. I risultati mostrano che ARCTIC supera i metodi di riferimento considerati, raggiungendo fino al 97,6% di allineamenti corretti sui dati simulati secondo il nostro criterio di valutazione. Sulle 36 misurazioni reali, ARCTIC stima correttamente tutte le trasformazioni, con errori medi di orientamento e di distanza inter-array rispettivamente pari a 4,77° e 9,5 cm. Questi risultati dimostrano il potenziale dello sfruttamento della riverberazione e delle informazioni specifiche del dominio acustico per l’autolocalizzazione di array di microfoni a partire da un’unica sorgente acustica.
Self-localization of microphone arrays in reverberant environments using the image source method
Guffroy, Thomas
2025/2026
Abstract
Self-localization of acoustic sensor networks is an important prerequisite for spatial audio applications, but remains challenging when the relative geometry of the network is unknown. This thesis addresses the problem of estimating the full 6-degree-of-freedom relative transformation between two synchronized, non-coplanar microphone arrays from a single acoustic source, without prior knowledge of the room geometry. The proposed approach exploits early acoustic reflections through the image source method, representing them as a 3D point cloud of virtual image sources. The task then amounts to matching and aligning the two point clouds estimated from the arrays, which is made difficult by their noisiness, the presence of spurious points, and the local symmetries produced by the room geometry. To this aim, we introduce ARCTIC (Anchor-based Registration using Congruent Triangle Consensus), a global registration algorithm specifically designed for acoustic image source point clouds. ARCTIC exploits the direct source as a natural correspondence, using it as an anchor to estimate the relative transformation. The method is evaluated on both simulated and real data. The results show that ARCTIC outperforms the considered baselines, reaching up to 97.6% of correct alignments on simulated data according to our evaluation criterion. Evaluated on 36 real measurements, ARCTIC correctly estimates all the transformations, with mean orientation and inter-array distance errors of 4.77° and 9.5 cm, respectively. These results demonstrate the potential of exploiting reverberation and domain-specific acoustic information for microphone-array self-localization from a single acoustic source.| File | Dimensione | Formato | |
|---|---|---|---|
|
Executive_summary_Thomas_GUFFROY.pdf
accessibile in internet per tutti a partire dal 23/09/2027
Descrizione: Executive Summary
Dimensione
831.41 kB
Formato
Adobe PDF
|
831.41 kB | Adobe PDF | Visualizza/Apri |
|
Thesis_report_Thomas_GUFFROY.pdf
accessibile in internet per tutti a partire dal 23/09/2027
Descrizione: Thesis Report
Dimensione
4.15 MB
Formato
Adobe PDF
|
4.15 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/10589/266795