Underwater SLAM using sonar images

Samec, Jaša (author), Skočaj, Danijel (mentor), Dobrevski, Matej (comentor)

Underwater vehicles have no access to GNSS, so a vehicle below the surface must estimate its trajectory by composing relative motion increments. Optical cameras fail in turbid water, which leaves the forward-looking imaging sonar as the most practical sensor for perception at long range. Sonar imagery is noisy, artifact-heavy, and weakly textured, so the pairwise keypoint matching on which existing odometry methods rely is fragile. We present ISOPoT, an imaging sonar odometry method that builds correspondences from multi-frame point tracks produced by a modern tracking-any-point model and converts them into a single globally consistent planar motion estimate through grid-based point management and correlation-based refinement. On the public Aracati 2017 dataset and on Portorož 2025, a dataset collected for this work, ISOPoT attains lower translation and rotation error than prior sonar odometry baselines in both the sonar-only and auxiliary-sensor (dead reckoning and magnetometer) settings. Finally, we close loops from sonar alone. An appearance-based detector proposes revisits and a learned relative-pose model turns them into pose-graph constraints, and neither stage consults the trajectory estimate, so detection does not fail precisely when the drift has grown large enough to require correction. On both Portorož 2025 sequences the accepted closure reduces absolute trajectory error with no auxiliary sensor.

• Language: English
• Keywords: Underwater robotics, Imaging sonar, Odometry, Point tracking, Motion estimation, Loop closure, SLAM
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-187932
• Publication date in RUL: 16.09.2026

Secondary language

• Language: Slovenian
• Title: Podvodni SLAM z uporabo sonarskih slik

Abstract

Podvodna vozila nimajo dostopa do GNSS, zato mora vozilo pod gladino svojo trajektorijo ocenjevati s sestavljanjem zaporednih ocen relativnega gibanja. Optične kamere v kalni vodi odpovejo, zato je naprej usmerjeni slikovni sonar najbolj praktičen senzor za zaznavanje na daljših razdaljah. Sonarske slike so šumne, vsebujejo artefakte in imajo šibko teksturo, zato je ujemanje značilnih točk med pari slik, na katerem temeljijo obstoječe metode odometrije, krhko. Predstavljamo ISOPoT, metodo odometrije slikovnega sonarja, ki korespondence gradi iz sledi točk čez več zaporednih slik, pridobljenih s sodobnim modelom za sledenje poljubnim točkam, in jih z mrežnim upravljanjem točk ter izboljšavo na podlagi korelacije pretvori v enotno, globalno konsistentno oceno ravninskega gibanja. Na javno dostopni podatkovni zbirki Aracati 2017 in na zbirki Portorož 2025, zajeti v okviru tega dela, ISOPoT dosega nižjo napako translacije in rotacije od prejšnjih metod sonarske odometrije, tako v načinu zgolj s sonarjem kot v načinu s pomožnimi senzorji (računska navigacija in magnetometer). Nazadnje zanke zapremo zgolj s sonarjem. Ponovne obiske predlaga detektor, ki temelji na videzu, naučen model relativne poze pa jih pretvori v vezi v grafu poz. Nobena od obeh stopenj ne uporablja ocene trajektorije, zato zaznava ne odpove ravno takrat, ko je napaka že dovolj velika, da bi jo bilo treba popraviti. Na obeh zaporedjih zbirke Portorož 2025 sprejeta zanka zmanjša absolutno napako trajektorije, in to brez pomožnih senzorjev.

• Keywords: Podvodna robotika, Slikovni sonar, Odometrija, Sledenje točkam, Ocenjevanje gibanja, Zapiranje zank, SLAM