Underwater navigation is a key technology for exploring the oceans and exploiting their resources. For autonomous underwater vehicles (AUVs) to explore the marine environment efficiently and securely, underwater simultaneous localization and mapping (SLAM) systems are often indispensable due to the lack of the global positioning system (GPS). In an underwater SLAM system, an AUV maps its surroundings and estimates its own pose at the same time. The pose of the AUV can be predicted by dead reckoning, but navigation errors accumulate over time. Therefore, sensors are needed to calibrate the state of the AUV. Among various sensors, the multibeam echosounder (MBES) is one of the most popular ones for underwater SLAM since it can acquire bathymetric point clouds with depth information of the surroundings. However, there are difficulties in data association for seabeds without distinct landmarks. Previous studies have focused more on traditional computer vision methods, which have limited performance on bathymetric data. In this thesis, a novel method based on deep learning is proposed to facilitate underwater perception. We conduct two experiments on place recognition and point cloud registration using data collected during a survey. The results show that, compared with the traditional methods, the proposed neural network is able to detect loop closures and register point clouds more efficiently. This work provides a better data association solution for designing underwater SLAM systems. / Undervattensnavigering är en viktig teknik för att utforska haven och utnyttja deras resurser. För att autonoma undervattensfordon (AUV) ska kunna utforska havsmiljön effektivt och säkert är underwater simultaneous localization and mapping (SLAM) system ofta oumbärliga på grund av bristen av det globala positioneringssystemet (GPS). I ett undervattens SLAM-system kartlägger ett AUV sin omgivning och uppskattar samtidigt sin egen position. AUV:s position kan förutsägas med hjälp av dödräkning, men navigeringsfel ackumuleras med tiden. Därför behövs sensorer för att kalibrera AUV:s tillstånd. Bland olika sensorer är multibeam ekolod (MBES) en av de mest populära för undervattens-SLAM eftersom den kan samla in batymetriska punktmoln med djupinformation om omgivningen. Det finns dock svårigheter med dataassociation för havsbottnar utan tydliga landmärken. Tidigare studier har fokuserat mer på traditionella datorvisionsmetoder som har begränsad prestanda för batymetriska data. I den här avhandlingen föreslås en ny metod baserad på djup inlärning för att underlätta undervattensuppfattning. Vi genomför två experiment på punktmolnregistrering med hjälp av data som samlats in under en undersökning. Resultaten visar att jämfört med de traditionella metoderna kan det föreslagna neurala nätverket upptäcka slingförslutningar och registrera punktmoln mer effektivt. Detta arbete ger en bättre lösning för dataassociation för utformning av undervattens SLAM-system.

L'imagerie numérique, de la synthèse d'images à la vision par ordinateur est en train de connaître une forte évolution, due entre autres facteurs à la démocratisation et au succès commercial des caméras 3D. Dans le même contexte, l'impression 3D grand public, qui est en train de vivre un essor fulgurant, contribue à la forte demande sur ce type de caméra pour les besoins de la numérisation 3D. L'objectif de cette thèse est d'acquérir et de maîtriser un savoir-faire dans le domaine de la capture/acquisition de modèles 3D en particulier sur l'aspect rendu réaliste. La réalisation d'un scanner 3D à partir d'une caméra RGB-D fait partie de l'objectif. Lors de la phase d'acquisition, en particulier pour un dispositif portable, on est confronté à deux problèmes principaux, le problème lié au référentiel de chaque capture et le rendu final de l'objet reconstruit. / Digital imaging, from the synthesis of images to computer vision isexperiencing a strong evolution, due among other factors to the democratization and commercial success of 3D cameras. In the same context, the consumer 3D printing, which is experiencing a rapid rise, contributes to the strong demand for this type of camera for the needs of 3D scanning. The objective of this thesis is to acquire and master a know-how in the field of the capture / acquisition of 3D models in particular on the rendered aspect. The realization of a 3D scanner from a RGB-D camera is part of the goal. During the acquisition phase, especially for a portable device, there are two main problems, the problem related to the repository of each capture and the final rendering of the reconstructed object.

When a Time-of-Flight (ToF) depth camera is used to monitor a region of interest, it has to be mounted correctly and have information regarding its position. Manual configuration currently require managing captured 3D ToF data in a 2D environment, which limits the user and might give rise to errors due to misinterpretation of the data. This thesis investigates if a real time 3D reconstruction mesh from a Microsoft HoloLens can be used as a target for point cloud registration using the ToF data, thus configuring the camera autonomously. Three registration algorithms, Fast Global Registration (FGR), Joint Registration Multiple Point Clouds (JR-MPC) and Prerejective RANSAC, were evaluated for this purpose. It was concluded that despite using different sensors it is possible to perform accurate registration. Also, it was shown that the registration can be done accurately within a reasonable time, compared with the inherent time to perform 3D reconstruction on the Hololens. All algorithms could solve the problem, but it was concluded that FGR provided the most satisfying results, though requiring several constraints on the data.

As the demand for 3D maps from LIDAR scanners increases, delivering high-quality maps becomes critical. One way to ensure the quality of such maps is through point cloud alignment classification, which aims to classify the alignment error between two registered point clouds. Specifically, we present the classifier FACT (Feature-Aware Classification Transformer), consisting of two main modules: feature extraction and classification. Descriptive features are extracted from the joint point cloud, which are then processed by a point transformer-based neural network to predict the alignment error class. In a ten-class point cloud alignment classification test, FACT achieved 92.4% accuracy, where the alignment error ranged from zero meters and radians to 0.9 meters and 0.09 radians. Remarkably, the classifier only made one misprediction beyond neighboring classes, exhibiting its ability to detect alignment errors as the classes have an inherent order. Furthermore, when benchmarked on two binary classification tasks, FACT showed significantly superior performance over the baseline and even obtained 100.0% accuracy for the easier of the two tasks. FACT not only detects potential errors in 3D maps but also estimates their magnitude, leading to more reliable 3D maps with quality estimations for each transformation.