Vehicle operations in underwater environments are frequently compromised by poor visibility conditions. The perception range of optical devices is heavily constrained in turbid waters, thus often complicating navigation and mapping tasks in environments such as harbors, bays, or rivers. A new generation of high-frequency forward-looking sonars that provide acoustic imagery at near-video frame rates have recently emerged as a promising alternative for working under these challenging conditions.<\/p>\n
In this thesis, we propose an end-to-end mosaicing framework tailored to the characteristics of forward-looking sonar imagery in order to build consistent overviews of planar underwater areas regardless of water visibility. Our solution targets versatility: it enables the generation of acoustic mosaics that involve roto-translational motions and comprise different vehicle tracklines; it is suitable for a wide range of scenarios, from feature-rich areas to environments with scarcity of features; it can be applicable on data collected with minimally instrumented vehicles; and it allows both offl\u00adine and real-time operation.<\/p>\n
The first problem to address is the pairwise registration of sonar images which is a key step in the mosaicing pipeline. The characteristics of the forward-looking sonar data, such as low and inhomogeneous resolution, low signal-to-noise ratio and intensity variations due to viewpoint changes, become a challenge for traditional feature-based registration techniques. For that reason we propose a Fourier-based methodology that, by involving all image content into the registration, offers robustness to noise and the different artifacts associated with the acoustic image formation. The approach relies on the phase correlation principle to estimate the image shifts and it is further adapted to cope with the multiple noise sources that can influence the registration, by providing specific masking, frequency filtering and rotation estimation procedures. When quantitatively compared, the proposed registration method shows superior performance to state-of-the-art feature-based approaches, while offering at the same time the possibility to be implemented efficiently.<\/p>\n
Next, we address the global alignment of the mosaic in order to enforce consistency between consecutive and non-consecutive image pairs. We lay out the problem as a graph optimization over the image poses, integrating spatial constraints from pairwise registrations as well as from navigation data when available. We provide a front-end to determine the constraints that should be included in the graph according to an initial estimation of the trajectory and a selection of potential overlapping candidate pairs. The work\u2011flow followed to build the graph is provided for both offl\u00adine mosaicing and the online approach, where constraints are added incrementally and under stringent restrictions to warrant real-time operation. In addition, we propose an uncertainty measure derived from the registration method to weigh appropriately the contribution of the registration constraints in the optimization.<\/p>\n
Finally, we explore the blending of the acoustic images into a smooth and informative mosaic while improving the signal-to-noise ratio and resolution of the final composition with respect to the individual frames. Furthermore, we identify the different photometric irregularities that can arise from the sonar imaging configuration and provide a set of strategies to minimize their impact both at frame and mosaic level.<\/p>\n
To validate the full proposed pipeline, an extensive experimental section is reported showing successful results in relevant field applications such as ship-hull inspection, harbor mapping and archaeological exploration.<\/p>\n
____________________________<\/p>\n
Freq\u00fcentment, les operacions amb vehicles en entorns submarins estan condicionades a la visibilitat de l\u2019aigua. El rang de percepci\u00f3 dels sensors \u00f2ptics \u00e9s molt limitat en aig\u00fces t\u00e8rboles i sovint fa complicades les tasques de navegaci\u00f3 i mapeig en entorns com ara ports, badies o rius. Recentment han aparegut al mercat una nova generaci\u00f3 de sonars de visi\u00f3 frontal, d\u2019alta freq\u00fc\u00e8ncia i amb la capacitat de generar imatges ac\u00fastiques a un alt ritme de refresc, convertint-se aix\u00ed en una alternativa prometedora per a operar en aquestes dif\u00edcils condicions.<\/p>\n
En aquesta tesi, proposem un sistema complet per a la construcci\u00f3 de mapes subaqu\u00e0tics adaptat a les caracter\u00edstiques d\u2019aquest tipus de sonars, per tal de construir mosaics d\u2019imatges ac\u00fastiques independentment de la visibilitat de l\u2019aigua. La soluci\u00f3 proposada t\u00e9 la versatilitat com a eix central: permet la generaci\u00f3 de mosaics ac\u00fastics que involucren moviments roto-translacionals i m\u00faltiples transectes, \u00e9s adequada per un ampli ventall d\u2019escenaris, des d\u2019\u00e0rees amb alta pres\u00e8ncia de punts d\u2019inter\u00e8s fins a entorns mancats de caracter\u00edstiques significants, \u00e9s aplicable a dades adquirides des de vehicles equipats amb poca instrumentaci\u00f3 i permet la generaci\u00f3 de mosaics tant en mode post-processat com en temps real.<\/p>\n
El primer problema que s\u2019adre\u00e7a \u00e9s el registre d\u2019imatges sonar, que \u00e9s un pas clau en el sistema de construcci\u00f3 de mosaics. Les caracter\u00edstiques de les imatges dels sonars de visi\u00f3 frontal, com ara la baixa resoluci\u00f3, la baixa relaci\u00f3 senyal-soroll, o les variacions d\u2019intensitat a causa dels canvis de punt de vista, causen moltes complicacions a les t\u00e8cniques que t\u00edpicament s\u2019utilitzen per registrar imatges \u00f2ptiques. Per aquesta ra\u00f3 proposem un m\u00e8tode basat en el domini freq\u00fcencial, el qual t\u00e9 en compte tot el contingut de les imatges en el proc\u00e9s de registre aconseguint aix\u00ed m\u00e9s robustesa davant del soroll i dels diferents artefactes associats amb la formaci\u00f3 de la imatge ac\u00fastica. El m\u00e8tode es basa en el principi de la correlaci\u00f3 de fase per computar els despla\u00e7aments de la imatge i s\u2019ha adaptat per fer front a les m\u00faltiples fonts de soroll que poden influenciar el registre proporcionant procediments espec\u00edfics d\u2019emmascarament, filtratge freq\u00fcencial i estimaci\u00f3 de la rotaci\u00f3. Comparacions quantitatives demostren que el m\u00e8tode proposat t\u00e9 un rendiment superior a altres t\u00e8cniques de registre de l\u2019estat de l\u2019art basades en punts d\u2019inter\u00e8s i ofereix, al mateix temps, la possibilitat de ser implementat eficientment.<\/p>\n
A continuaci\u00f3 s\u2019ha adre\u00e7at l\u2019alineament global del mosaic per tal d\u2019imposar consist\u00e8ncia entre parelles d\u2019imatges consecutives i no consecutives. S\u2019ha plantejat el problema com la optimitzaci\u00f3 d\u2019un graf on els nodes s\u00f3n les posicions de les imatges i les restriccions espacials entre ells provenen del registre de les diferents parelles o de les dades de navegaci\u00f3, en cas de disposar d\u2019aquestes. Detallem l\u2019estrat\u00e8gia a seguir per determinar quines restriccions s\u2019han d\u2019incorporar al graf d\u2019acord amb una estimaci\u00f3 inicial de la traject\u00f2ria del vehicle i una selecci\u00f3 de les parelles d\u2019imatges que poden tenir solapament. El procediment per a la construcci\u00f3 del graf es descriu tant per l\u2019elaboraci\u00f3 de mosaics offl\u00adine<\/i> com online<\/i>. En aquest darrer cas, les restriccions s\u2019afegeixen incrementalment i sota estrictes l\u00edmits per garantir l\u2019execuci\u00f3 en temps real. A m\u00e9s a m\u00e9s, proposem una mesura d\u2019incertesa derivada del m\u00e8tode de registre per tal de donar el pes adequat a cada restricci\u00f3 dins de la optimitzaci\u00f3.<\/p>\n
Finalment, s\u2019explora la fusi\u00f3 de les imatges ac\u00fastiques en un \u00fanic mosaic d\u2019aparen\u00e7a n\u00edtida i informativa, aconseguint al mateix temps una relaci\u00f3 senyal-soroll i una resoluci\u00f3 millorades respecte les de les imatges individuals. Tamb\u00e9 s\u2019identifiquen les diferents irregularitats fotom\u00e8triques que poden sorgir degut a la configuraci\u00f3 i col\u00b7locaci\u00f3 del sonar en l\u2019entorn i es proporcionen un seguit d\u2019estrat\u00e8gies per minimitzar el seu impacte tant a nivell d\u2019imatge com a nivell de mosaic.<\/p>\n
Per tal d\u2019avaluar la totalitat del marc de treball proposat, es presenta una extensiva secci\u00f3 experimental demostrant resultats satisfactoris en aplicacions rellevants com ara inspecci\u00f3 de cascos de vaixells, mapeig de ports o exploracions arqueol\u00f2giques.<\/p>\n","protected":false},"excerpt":{"rendered":"