Submitted by JOVANNY BEDOYA GUAPACHA null (jovan@utp.edu.co) on 2017-11-02T14:40:57Z No. of bitstreams: 1 TESE _JBG_verf__02_11_2017_repositorio.pdf: 4463035 bytes, checksum: a4e99464884d8580fc971b9f062337d4 (MD5) / Approved for entry into archive by LUIZA DE MENEZES ROMANETTO (luizamenezes@reitoria.unesp.br) on 2017-11-13T16:46:44Z (GMT) No. of bitstreams: 1 guapacha_jb_dr_ilha.pdf: 4463035 bytes, checksum: a4e99464884d8580fc971b9f062337d4 (MD5) / Made available in DSpace on 2017-11-13T16:46:44Z (GMT). No. of bitstreams: 1 guapacha_jb_dr_ilha.pdf: 4463035 bytes, checksum: a4e99464884d8580fc971b9f062337d4 (MD5) Previous issue date: 2017-09-04 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES) / A criação de robôs que podem operar autonomamente em ambientes controlados e não controlados tem sido, um dos principais objetivos da robótica móvel. Para que um robô possa navegar em um ambiente interno desconhecido, ele deve se localizar e ao mesmo tempo construir um mapa do ambiente que o rodeia, a este problema dá-se o nome de Localização e Mapeamento Simultâneos- SLAM. Tem-se como proposta neste trabalho para solucionar o problema do SLAM, o uso de um sensor RGB-D, com 6 graus de liberdade para perceber o ambiente, o qual é embarcado em um robô. O problema do SLAM pode ser solucionado estimando a pose - posição e orientação, e a trajetória do sensor no ambiente, de forma precisa, justificando a construção de um mapa em três dimensões (3D). Esta estimação envolve a captura consecutiva de frames do ambiente fornecidos pelo sensor RGB-D, onde são determinados os pontos mais acentuados das imagens através do uso de características visuais dadas pelo algoritmo ORB. Em seguida, a comparação entre frames consecutivos e o cálculo das transformações geométricas são realizadas, mediante o algoritmo de eliminação de correspondências atípicas, bPROSAC. Por fim, uma correção de inconsistências é efetuada para a reconstrução do mapa 3D e a estimação mais precisa da trajetória do robô, utilizando técnicas de otimização não lineares. Experimentos são realizados para mostrar a construção do mapa e o desempenho da proposta. / The robots creation that can operate autonomously in controlled and uncontrolled environments has been, one of the main objectives of mobile robotics. In order for a robot to navigate in an unknown internal environment, it must locate yourself and at the same time construct a map of the surrounding environment this problem is called Simultaneous Location and Mapping - SLAM. The purpose of this work for solution to SLAM’s problem is to use an RGB-D sensor with 6 degrees of freedom to perceive the environment, which is embedded onto a robot.The SLAM's problem can be solved by estimating the position and orientation, and the path of the sensor/robot in the environment, in precise form, justifying the construction of a 3D map. This estimation involves the consecutive capture of the environment's frames provided by the RGB-D sensor, where the pronounced points of the images are determined through the use of visual characteristics given by the ORB algorithm. Then, the comparison between consecutive frames and the calculation of the geometric transformations are performed using the algorithm of elimination of atypical correspondences, bPROSAC. Finally, a correction of inconsistencies is made for the reconstruction of the 3D map and the more accurate estimation of the robot trajectory, using non-linear optimization techniques. Experiments are carried out to show the construction of the map and the performance of the proposal.

Made available in DSpace on 2014-12-17T14:55:46Z (GMT). No. of bitstreams: 1 GuilhermeLS_DISSERT.pdf: 861871 bytes, checksum: 8461d130e59e8fb9ac951602b094fd18 (MD5) Previous issue date: 2010-05-07 / Coordena??o de Aperfei?oamento de Pessoal de N?vel Superior / The development and refinement of techniques that make simultaneous localization and mapping (SLAM) for an autonomous mobile robot and the building of local 3-D maps from a sequence of images, is widely studied in scientific circles. This work presents a monocular visual SLAM technique based on extended Kalman filter, which uses features found in a sequence of images using the SURF descriptor (Speeded Up Robust Features) and determines which features can be used as marks by a technique based on delayed initialization from 3-D straight lines. For this, only the coordinates of the features found in the image and the intrinsic and extrinsic camera parameters are avaliable. Its possible to determine the position of the marks only on the availability of information of depth. Tests have shown that during the route, the mobile robot detects the presence of characteristics in the images and through a proposed technique for delayed initialization of marks, adds new marks to the state vector of the extended Kalman filter (EKF), after estimating the depth of features. With the estimated position of the marks, it was possible to estimate the updated position of the robot at each step, obtaining good results that demonstrate the effectiveness of monocular visual SLAM system proposed in this paper / O desenvolvimento e aperfei?oamento de t?cnicas que fa?am simultaneamente o mapeamento e a localiza??o (Simultaneous Localization and Mapping - SLAM) de um rob? m?vel aut?nomo e a cria??o de mapas locais 3-D, a partir de uma sequ?ncia de imagens, ? bastante estudada no meio cient?fico. Neste trabalho ? apresentado uma t?cnica de SLAM visual monocular baseada no filtro de Kalman estendido, que utiliza caracter?sticas encontradas em uma sequ?ncia de imagens atrav?s do descritor SURF (Speeded Up Robust Features) e determina quais caracter?sticas podem ser utilizadas como marcas atrav?s de uma t?cnica de inicializa??o atrasada baseada em retas 3-D. Para isso, tem-se dispon?vel apenas as coordenadas das caracter?sticas detectadas na imagem e os par?metros intr?nsecos e extr?nsecos da c?mera. ? poss?vel determinar a posi??o das marcas somente com a disponibilidade da informa??o de profundidade. Os experimentos realizados mostraram que durante o percurso, o rob? m?vel detecta a presen?a de caracter?sticas nas imagens e, atrav?s de uma t?cnica proposta para inicializa??o atrasada de marcas, adiciona novas marcas ao vetor de estados do filtro de Kalman estendido (FKE) ap?s estimar a profundidade das caracter?sticas. Com a posi??o estimada das marcas, foi poss?vel estimar a posi??o atualizada do rob? a cada passo; obtendo resultados satisfat?rios que comprovam a efici?ncia do sistema de SLAM visual monocular proposto neste trabalho

Robôs verdadeiramente autônomos devem conhecer o ambiente para executar tarefas complexas. Em ambientes desconhecidos o robô deve concorrentemente construir o mapa do ambiente e se localizar usando sensores proprioceptivos e exteroceptivos imprecisos. Isto é problemático, uma vez que o mapa parcial e possivelmente incorreto do ambiente será usado para corrigir erros de localização. Este problema importante da robótica móvel é conhecido como Localização e Mapeamento Simultâneos (SLAM). Quando um robô autonomamente executa o algoritmo de SLAM concorrentemente com uma estratégia de exploração, o problema passa a se chamar SLAM Ativo ou Exploração Integrada. Um dos principais desafios por trás destes problemas é o tratamento de fechamento de ciclos. Ao atravessar regiões desconhecidas ou ambientes esparsos, a pose do robô e o mapa podem não ser propriamente corrigidos por falta de informação. Quando isto acontece, as incertezas da posição do robô e do mapa aumentam, podendo levar a erros irrecuperáveis. Por outro lado, quando o ciclo é fechado corretamente, estas incertezas diminuem consideravelmente. Portanto, a escolha do caminho para explorar o ambiente pode drasticamente melhorar ou degradar a qualidade do mapeamento e da localização. Uma técnica bem conhecida de exploração de ambientes é a adaptação do problema de valor de contorno (BVP) para a equação de Laplace e condições de contorno de Dirichlet. Apesar de ser fácil de implementar, resultando em trajetórias de exploração suaves, esta técnica não endereça cuidadosamente erros de SLAM, uma vez que ela segue a descida do gradiente, o que pode não possibilitar revisitas, uma limitação crucial para o SLAM Ativo. Mesmo sendo uma técnica de exploração gulosa e direcionada a fronteiras, consideramos que a flexibilidade do BVP e condições de contorno de Dirichlet ainda são pouco exploradas. Nossa proposta é modificar o algoritmo de Exploração por BVP para executar comportamentos complexos, tais como revisitas e, em particular, fechamentos de ciclo. Apresentamos duas novas abordagens: a primeira faz uso de uma condição de contorno direcionada pelo tempo combinada a distorções de potencial para gerar comportamentos de fechamento de ciclo, além de um potencial que nunca cessa de existir, mesmo após o ambiente ter sido completamente explorado; a segunda, propicia o fechamento de ciclos aproveitando a propagação do potencial em regiões desconhecidas, através de um par dinâmico de condições de contorno que funcionam como obstáculos e objetivos virtuais. Ambas abordagens aproveitam o Esqueleto de Voronoi do ambiente para reduzir o custo computacional do algoritmo. Testes em ambientes reais e simulados usando o robô Pioneer 3DX mostram que as técnicas apresentadas apresetam melhores resultados quando comparadas a técnicas concorrentes. / Truly autonomous robots must know the environment in order to execute complex tasks. In unknown environments, the robot must construct a map and localize itself using noisy proprioceptive and exteroceptive sensors. This is problematic, since the partial and possibly inaccurate map of the environment will be used to correct localization errors. This important problem of mobile robotics is known as Simultaneous Localization and Mapping (SLAM). When a robot autonomously execute a SLAM algorithm concurrently with an exploration strategy, this problem is called Active SLAM or Integrated Exploration. One of the main challenges behind both these problems is the treatment of loop closures. While the robot traverses unknown regions or sparse environments, the robot pose and the map may not be properly corrected due to lack of information. When this happens, the uncertainties about the map and the robot pose increase, which may lead to unrecoverable SLAM errors. On the other hand, when a loop is closed successfully, these uncertainties drastically decrease. Therefore, path chosen to explore the environment can considerably improve or degrade the quality of both localization and mapping. One well known way to explore the environment is the adaptation of the Boundary Value Problem (BVP) for the Laplace Equation and Dirichlet boundary conditions. Even though it is easy to implement, resulting in smooth exploration trajectories, it does not carefully address SLAM errors, since it follows a gradient decent which not always allows revisits, a crucial limitation for Active SLAM. Despite being a greedy frontier driven exploration strategy, we consider the flexibility of the BVP and Dirichlet boundary conditions still under-explored for Active SLAM. Our proposal is to modify the BVP Exploration algorithm to execute complex exploration behaviors, such as revisits and, in particular, loop-closures. We present two new approaches: the first makes use of a time driven boundary value condition together with potential distortions to generate loop closing behaviors and a potential field that never ceases to exist, even after the exploration ends; the second enables loop closure behaviors with BVP by taking advantage of potential propagation in unknown space generated by a pair of dynamic boundary conditions functioning as virtual walls and goals. Both approaches take advantage of a local optimization that uses the Voronoi Skeleton to reduce the computational cost of the algorithm. Tests in real and simulated environments using a Pioneer 3DX show that the proposed approaches present better results when compared with competing approaches.

Localização e Mapeamento Simultâneos (SLAM) é uma das tarefas mais difíceis em robótica móvel, uma vez que existe uma dependência mútua entre a estimativa da localização do robô e a construção do mapa de ambiente. As estratégias de SLAM mais bem sucedidas focam na construção de um mapa métrico probabilístico empregando técnicas de filtragem Bayesiana. Embora tais métodos permitam a construção de soluções localmente consistentes e coerentes, o SLAM continua sendo um problema crítico em operações em ambientes grandes. Para contornar esta limitação, muitas estratégias dividem o ambiente em pequenas regiões, e formulam o problema de SLAM como uma combinação de múltiplos submapas métricos precisos associados em um mapa topológico. Este trabalho propõe um método de SLAM baseado nos algoritmos DP-SLAM (Distributed Particle SLAM) e SegSlam (Segmented SLAM). SegSLAM é um algoritmo que cria múltiplos submapas para cada região do ambiente, e posteriormente constrói o mapa global selecionando combinações de submapas. Por sua vez, DP-SLAM é um algoritmo de filtro de particulas Rao-Blackwellized que utiliza uma representação distribuída eficiente dos mapas das partículas, juntamente com a árvore de ascendência das partículas. A característica distribuída destas estruturas é favorável para a combinação de diferentes segmentos de mapa localmente precisos, o que aumenta a diversidade de soluções. O algoritmo proposto nesta dissertação, chamado SDP-SLAM, segmenta e combina diferentes hipóteses de trajetórias do robô, a fim de reconstruir o mapa do ambiente. Nossas principais contribuições são o desenvolvimento de novas estratégias para o casamento de submapas e para a estimativa de boas combinações de submapas. O SDP-SLAM foi avaliado através de experimentos realizados por um robô móvel operando em ambientes reais e simulados. / Simultaneous Localization and Mapping (SLAM) is one of the most difficult tasks in mobile robotics, since there is a mutual dependency between the estimation of the robot pose and the construction of the environment map. Most successful strategies in SLAM focus in building a probabilistic metric map employing Bayesian filtering techniques. While these methods allow the construction of consistent and coherent local solutions, the SLAM remains a critical problem in operations within large environments. To circumvent this limitation, many strategies divide the environment in small regions, and formulate the SLAM problem as a combination of multiple precise metric submaps associated in a topological map. This work proposes a SLAM method based on the Distributed Particle SLAM (DPSLAM) and the Segmented SLAM (SegSLAM) algorithms. SegSLAM is an algorithm that generates multiple submaps for every region of the environment, and then build the global map by selecting combinations of submaps. DP-SLAM is a Rao-Blackwellized particle filter algorithm that uses an efficient distributed representation of the particles maps associated with an ancestry tree of the particles. The distributed characteristic of these structures favors the combination of locally accurate map segments, that can increase the diversity of global level solutions. The algorithm proposed in this dissertation, called SDP-SLAM, segments and combines different hypotheses of robot trajectories to reconstruct the environment map. Our main contributions are the development of novel strategies for the matching of submaps and for the estimation of good submaps combinations. SDP-SLAM was evaluated through experiments performed by a mobile robot operating in real and simulated environments.

La thématique de la navigation autonome constitue l’un des principaux axes de recherche dans le domaine des véhicules intelligents et des robots mobiles. Dans ce contexte, on cherche à doter le robot d’algorithmes et de méthodes lui permettant d’évoluer dans un environnement complexe et dynamique, en toute sécurité et en parfaite autonomie. Dans ce contexte, les algorithmes de localisation et de cartographie occupent une place importante. En effet, sans informations suffisantes sur la position du robot (localisation) et sur la nature de son environnement (cartographie), les autres algorithmes (génération de trajectoire, évitement d’obstacles ...) ne peuvent pas fonctionner correctement. Nous avons centré notre travail de thèse sur une problématique précise : développer un algorithme de SLAM simple, rapide, léger et limitant les erreurs de localisation et de cartographie au maximum sans fermeture de boucle. Au cœur de notre approche, on trouve un algorithme d’IML : Incremental Maximum Likelihood. Ce type d’algorithmes se base sur une estimation itérative de la localisation et de la cartographie. Il est ainsi naturellement divergent. Le choix de l’IML est justifié essentiellement par sa simplicité et sa légèreté. La particularité des travaux réalisés durant cette thèse réside dans les différents outils et algorithmes utilisés afin de limiter la divergence de l’IML au maximum, tout en conservant ses avantages. / One of the main areas of research in the field of intelligent vehicles and mobile robots is Autonomous navigation. In this field, we seek to create algorithms and methods that give robots the ability to move safely and autonomously in a complex and dynamic environment. In this field, localization and mapping algorithms have an important place. Indeed,without reliable information about the robot position (localization) and the nature of its environment (mapping), the other algorithms (trajectory generation, obstacle avoidance ...) cannot achieve their tasks properly. We focused our work during this thesis on a specific problem: to develop a simple, fast and lightweight SLAM algorithm that can minimize localization errors without loop closing. At the center of our approach, there is an IML algorithm: Incremental Maximum Likelihood. This kind of algorithms is based on an iterative estimation of the localization and the mapping. It contains thus naturally a growing error in the localization process. The choice of IML isjustified mainly by its simplicity and lightness. The main idea of our work is built around thedifferent tools and algorithms used to minimize the localization error of IML, while keeping its advantages.

Localização e Mapeamento são problemas fundamentais da robótica que vêm sendo estudados exaustivamente pela comunidade científica para a navegação de robôs móveis. A maior parte das pesquisas estão concentradas em implementações para computadores pessoais, mas pouco se tem feito na área de computação embarcada. Este trabalho mostra a análise e implementação em FPGA de um algoritmo de localização para ambientes dinâmicos composto por um filtro de partículas, e também de um algoritmo de mapeamento baseado na técnica de scan matching. Os algoritmos originais desenvolvidos em linguagem de programação C foram analisados e modificados para uma abordagem embarcada (embedded) em robôs reconfiguráveis utilizando-se o processador Nios II da Altera. Os algoritmos são comparados quanto ao desempenho, no intuito de servir como referência no futuro desenvolvimento da ferramenta de codesign autom´atico ARCHITECT+ / Localization and Mapping are fundamental robot navigation problems wich currently has been exaustily studied by scientific comunity. Most of research is concentrated on implementation for personal computers, and the robot navigation is done on static environment. But, these algorithms can not be directly applied for embedded solutions on dynamic environments. This work shows an analysis and implementation on FPGA of a localization algorithm for dynamic environments composed of a particle filter, and by an mapping algorthm known as scan matching. The original algorithm devoloped on C programming language for PCs are analised and modified for an embedded approach to mobile robots using Altera Nios II processor. Both C and embedded algorithms are compared within performance, to serve as reference on a future developement of automatic codesign tool ARCHITECT+

Dans cette thèse, nous présentons une nouvelle méthode de cartographie visuelle hybride qui exploite des informations métriques, topologiques et sémantiques. Notre but est de réduire le coût calculatoire par rapport à des techniques de cartographie purement métriques. Comparé à de la cartographie topologique, nous voulons plus de précision ainsi que la possibilité d’utiliser la carte pour le guidage de robots. Cette méthode hybride de construction de carte comprend deux étapes. La première étape peut être vue comme une carte topo-métrique avec des nœuds correspondants à certaines régions de l’environnement. Ces cartes sont ensuite complétées avec des données métriques aux nœuds correspondant à des sous-séquences d’images acquises quand le robot revenait dans des zones préalablement visitées. La deuxième étape augmente ce modèle en ajoutant des informations sémantiques. Une classification est effectuée sur la base des informations métriques en utilisant des champs de Markov conditionnels (CRF) pour donner un label sémantique à la trajectoire locale du robot (la route dans notre cas) qui peut être "doit", "virage" ou "intersection". L’information métrique des secteurs de route en virage ou en intersection est conservée alors que la métrique des lignes droites est effacée de la carte finale. La fermeture de boucle n’est réalisée que dans les intersections ce qui accroît l’efficacité du calcul et la précision de la carte. En intégrant tous ces nouveaux algorithmes, cette méthode hybride est robuste et peut être étendue à des environnements de grande taille. Elle peut être utilisée pour la navigation d’un robot mobile ou d’un véhicule autonome en environnement urbain. Nous présentons des résultats expérimentaux obtenus sur des jeux de données publics acquis en milieu urbain pour démontrer l’efficacité de l’approche proposée. / In this thesis, a novel vision based hybrid mapping framework which exploits metric, topological and semantic information is presented. We aim to obtain better computational efficiency than pure metrical mapping techniques, better accuracy as well as usability for robot guidance compared to the topological mapping. A crucial step of any mapping system is the loop closure detection which is the ability of knowing if the robot is revisiting a previously mapped area. Therefore, we first propose a hierarchical loop closure detection framework which also constructs the global topological structure of our hybrid map. Using this loop closure detection module, a hybrid mapping framework is proposed in two step. The first step can be understood as a topo-metric map with nodes corresponding to certain regions in the environment. Each node in turn is made up of a set of images acquired in that region. These maps are further augmented with metric information at those nodes which correspond to image sub-sequences acquired while the robot is revisiting the previously mapped area. The second step augments this model by using road semantics. A Conditional Random Field based classification on the metric reconstruction is used to semantically label the local robot path (road in our case) as straight, curved or junctions. Metric information of regions with curved roads and junctions is retained while that of other regions is discarded in the final map. Loop closure is performed only on junctions thereby increasing the efficiency and also accuracy of the map. By incorporating all of these new algorithms, the hybrid framework presented can perform as a robust, scalable SLAM approach, or act as a main part of a navigation tool which could be used on a mobile robot or an autonomous car in outdoor urban environments. Experimental results obtained on public datasets acquired in challenging urban environments are provided to demonstrate our approach.

Il est fondamental pour un robot d'être capable de se déplacer de manière complètement autonome afin d'accomplir une mission qui lui a été confiée, et ce avec un budget énergétique fini, dans un laps de temps contraint et sans connaissances préalables de l’environnement. Afin d'atteindre un objectif dans le plan ou l'espace, un robot doit à minima être capable d'accomplir quatre tâches: maintenir une représentation abstraite de l'environnement (une carte), être capable de se localiser à l'intérieur de cette représentation, utiliser la représentation pour planifier des itinéraires et naviguer le long de la trajectoire prévue tout en s'adaptant aux dynamiques de l'environnement et en évitant les obstacles. Chacun de ces problèmes a été étudié par la communauté de la robotique. Cependant, ces quatre composants sont en général étudiés séparément et sont par conséquent incompatibles entre eux pour l'essentiel. De plus, étant donné qu'humains et robots ne disposent que de ressources computationelles et mémorielles finies, les algorithmes de planification, navigation et SLAM devraient être capables de fonctionner avec des données incomplètes ou compressées tout en garantissant que le ou les objectifs fixés soient atteints. Dans cette thèse, la planification, la navigation et le SLAM dans des environnements arbitrairement grands et avec des ressources computationelles et mémorielles finies sont vues comme un seul problème, créant un nouveau paradigme que nous appelons Navigation Exploratoire au long de la Vie ou Lifelong Exploratory Navigation. / One of the yet unresolved canonical problems of robotics is to have robots move completely autonomously in order to accomplish any mission they are charged with, with time and resource constraints and without prior knowledge of the environment. Reaching a goal requires the robot to perform at least four tasks: maintaining an abstract representation of the environment (map), being able to localize itself within this representation, using the representation to plan paths and navigating on the planned paths while handling dynamics of the environment and avoiding obstacles. Each of these problems has been studied extensively by the robotics community. However, the four components are usually studied separately, and as a result are mostly incompatible with each other. Additionally, since humans as well as robots have to operate with finite memory and computing resources, long running planning, navigation and SLAM algorithms may have to operate on incomplete or compressed data while guaranteeing that the goal(s) can still be reached. In this thesis, planning, navigation and SLAM in arbitrarily large environments with finite computing resources and memory are considered as one single problem, for a new bio-inspired paradigm which we call Lifelong Exploratory Navigation.

Tato diplomová práce se zabývá tvorbou fúze pozičních dat z existující realtimové im- plementace vizuálního SLAMu a kolové odometrie. Výsledkem spojení dat je potlačení nežádoucích chyb u každé ze zmíněných metod měření, díky čemuž je možné vytvořit přesnější 3D model zkoumaného prostředí. Práce nejprve uvádí teorií potřebnou pro zvládnutí problematiky 3D SLAMu. Dále popisuje vlastnosti použitého open source SLAM projektu a jeho jednotlivé softwarové úpravy. Následně popisuje principy spo- jení pozičních informací získaných vizuálními a odometrickými snímači, dále uvádí popis diferenciálního podvozku, který byl použit pro tvorbu kolové odometrie. Na závěr práce shrnuje výsledky dosažené datovou fúzí a srovnává je s původní přesností vizuálního SLAMu.

Cette thèse s'intéresse au problème du positionnement (position et orientation) dans un contexte de réalité augmentée et aborde spécifiquement les solutions à base de capteurs embarqués. Aujourd'hui, les systèmes de navigation vision-inertiel commencent à combler les besoins spécifiques de cette application. Néanmoins, ces systèmes se basent tous sur des corrections de trajectoire issues des informations visuelles à haute fréquence afin de pallier la rapide dérive des capteurs inertiels bas-coûts. Pour cette raison, ces méthodes sont mises en défaut lorsque l'environnement visuel est défavorable.Parallèlement, des travaux récents menés par la société Sysnav ont démontré qu'il était possible de réduire la dérive de l'intégration inertielle en exploitant le champ magnétique, grâce à un nouveau type d'UMI bas-coût composée – en plus des accéléromètres et gyromètres traditionnels – d'un réseau de magnétomètres. Néanmoins, cette méthode est également mise en défaut si des hypothèses de non-uniformité et de stationnarité du champ magnétique ne sont pas vérifiées localement autour du capteur.Nos travaux portent sur le développement d'une solution de navigation à l'estime robuste combinant toutes ces sources d'information: magnétiques, visuelles et inertielles.Nous présentons plusieurs approches pour la fusion de ces données, basées sur des méthodes de filtrage ou d’optimisation et nous développons un modèle de prédiction du champ magnétique inspiré d'approximation proposées en inertiel et permettant d’intégrer efficacement des termes magnétiques dans les méthodes d’ajustement de faisceaux. Les performances de ces différentes approches sont évaluées sur des données réelles et nous démontrons le bénéfice de la fusion de données comparées aux solutions vision-inertielles ou magnéto-inertielles. Des propriétés théoriques de ces méthodes liées à la théorie de l’invariance des estimateurs sont également étudiées. / This thesis addresses the issue of positioning in 6-DOF that arises from augmented reality applications and focuses on embedded sensors based solutions.Nowadays, the performance reached by visual-inertial navigation systems is starting to be adequate for AR applications. Nonetheless, those systems are based on position correction from visual sensors involved at a relatively high frequency to mitigate the quick drift of low-cost inertial sensors. This is a problem when the visual environment is unfavorable.In parallel, recent works have shown it was feasible to leverage magnetic field to reduce inertial integration drift thanks to a new type of low-cost sensor, which includes – in addition to the accelerometers and gyrometers – a network of magnetometers. Yet, this magnetic approach for dead-reckoning fails if stationarity and non-uniformity hypothesis on the magnetic field are unfulfilled in the vicinity of the sensor.We develop a robust dead-reckoning solution combining simultaneously information from all these sources: magnetic, visual, and inertial sensor. We present several approaches to solve for the fusion problem, using either filtering or non-linear optimization paradigm and we develop an efficient way to use magnetic error term in a classical bundle adjustment that was inspired from already used idea for inertial terms. We evaluate the performance of these estimators on data from real sensors. We demonstrate the benefits of the fusion compared to visual-inertial and magneto-inertial solutions. Finally, we study theoretical properties of the estimators that are linked to invariance theory.