The purpose of this thesis is to investigate the applicability of a certain model based classification algorithm. The algorithm is centered around a flexible wireframe prototype that can instantiate a number of different vehicle classes such as a hatchback, pickup or a bus to mention a few. The parameters of the model are fitted using Newton minimization of errors between model line segments and observed line segments. Furthermore a number of methods for object detection based on motion are described and evaluated. Results from both experimental and real world data is presented.

<p>The purpose of this thesis is to investigate the applicability of a certain model based classification algorithm. The algorithm is centered around a flexible wireframe prototype that can instantiate a number of different vehicle classes such as a hatchback, pickup or a bus to mention a few. The parameters of the model are fitted using Newton minimization of errors between model line segments and observed line segments. Furthermore a number of methods for object detection based on motion are described and evaluated. Results from both experimental and real world data is presented.</p>

This thesis addresses visual tracking of a non-cooperative as well as a partially cooperative satellite, to enable close-range rendezvous between a servicer and a target satellite. Visual tracking and estimation of relative motion between a servicer and a target satellite are critical abilities for rendezvous and proximity operation such as repairing and deorbiting. For this purpose, Lidar has been widely employed in cooperative rendezvous and docking missions. Despite its robustness to harsh space illumination, Lidar has high weight and rotating parts and consumes more power, thus undermines the stringent requirements of a satellite design. On the other hand, inexpensive on-board cameras can provide an effective solution, working at a wide range of distances. However, conditions of space lighting are particularly challenging for image based tracking algorithms, because of the direct sunlight exposure, and due to the glossy surface of the satellite that creates strong reflection and image saturation, which leads to difficulties in tracking procedures. In order to address these difficulties, the relevant literature is examined in the fields of computer vision, and satellite rendezvous and docking. Two classes of problems are identified and relevant solutions, implemented on a standard computer are provided. Firstly, in the absence of a geometric model of the satellite, the thesis presents a robust feature-based method with prediction capability in case of insufficient features, relying on a point-wise motion model. Secondly, we employ a robust model-based hierarchical position localization method to handle change of image features along a range of distances, and localize an attitude-controlled (partially cooperative) satellite. Moreover, the thesis presents a pose tracking method addressing ambiguities in edge-matching, and a pose detection algorithm based on appearance model learning. For the validation of the methods, real camera images and ground truth data, generated with a laboratory tet bed similar to space conditions are used. The experimental results indicate that camera based methods provide robust and accurate tracking for the approach of malfunctioning satellites in spite of the difficulties associated with specularities and direct sunlight. Also exceptional lighting conditions associated to the sun angle are discussed, aimed at achieving fully reliable localization system in a certain mission.

Human actions are the essence of a human life and a natural product of the human mind. Analysis of human activities by a machine has attracted the attention of many researchers. This analysis is very important in a variety of domains including surveillance, video retrieval, human-computer interaction, athlete performance investigation, etc. This dissertation makes three major contributions to automatic analysis of human actions. First, we conjecture that the relationship between body joints of two actors in the same posture can be described by a 3D rigid transformation. This transformation simultaneously captures different poses and various sizes and proportions. As a consequence of this conjecture, we show that there exists a fundamental matrix between the imaged positions of the body joints of two actors, if they are in the same posture. Second, we propose a novel projection model for cameras moving at a constant velocity in 3D space, \emph cameras, and derive the Galilean fundamental matrix and apply it to human action recognition. Third, we propose a novel use for the invariant ratio of areas under an affine transformation and utilizing the epipolar geometry between two cameras for 2D model-based tracking of human body joints. In the first part of the thesis, we propose an approach to match human actions using semantic correspondences between human bodies. These correspondences are used to provide geometric constraints between multiple anatomical landmarks ( e.g. hands, shoulders, and feet) to match actions observed from different viewpoints and performed at different rates by actors of differing anthropometric proportions. The fact that the human body has approximate anthropometric proportion allows for innovative use of the machinery of epipolar geometry to provide constraints for analyzing actions performed by people of different anthropometric sizes, while ensuring that changes in viewpoint do not affect matching. A novel measure in terms of rank of matrix constructed only from image measurements of the locations of anatomical landmarks is proposed to ensure that similar actions are accurately recognized. Finally, we describe how dynamic time warping can be used in conjunction with the proposed measure to match actions in the presence of nonlinear time warps. We demonstrate the versatility of our algorithm in a number of challenging sequences and applications including action synchronization , odd one out, following the leader, analyzing periodicity etc. Next, we extend the conventional model of image projection to video captured by a camera moving at constant velocity. We term such moving camera Galilean camera. To that end, we derive the spacetime projection and develop the corresponding epipolar geometry between two Galilean cameras. Both perspective imaging and linear pushbroom imaging form specializations of the proposed model and we show how six different ``fundamental" matrices including the classic fundamental matrix, the Linear Pushbroom (LP) fundamental matrix, and a fundamental matrix relating Epipolar Plane Images (EPIs) are related and can be directly recovered from a Galilean fundamental matrix. We provide linear algorithms for estimating the parameters of the the mapping between videos in the case of planar scenes. For applying fundamental matrix between Galilean cameras to human action recognition, we propose a measure that has two important properties. First property makes it possible to recognize similar actions, if their execution rates are linearly related. Second property allows recognizing actions in video captured by Galilean cameras. Thus, the proposed algorithm guarantees that actions can be correctly matched despite changes in view, execution rate, anthropometric proportions of the actor, and even if the camera moves with constant velocity. Finally, we also propose a novel 2D model based approach for tracking human body parts during articulated motion. The human body is modeled as a 2D stick figure of thirteen body joints and an action is considered as a sequence of these stick figures. Given the locations of these joints in every frame of a model video and the first frame of a test video, the joint locations are automatically estimated throughout the test video using two geometric constraints. First, invariance of the ratio of areas under an affine transformation is used for initial estimation of the joint locations in the test video. Second, the epipolar geometry between the two cameras is used to refine these estimates. Using these estimated joint locations, the tracking algorithm determines the exact location of each landmark in the test video using the foreground silhouettes. The novelty of the proposed approach lies in the geometric formulation of human action models, the combination of the two geometric constraints for body joints prediction, and the handling of deviations in anthropometry of individuals, viewpoints, execution rate, and style of performing action. The proposed approach does not require extensive training and can easily adapt to a wide variety of articulated actions.

Les travaux réalisés au cours de cette thèse s’inscrivent dans la problématique de localisation en temps réel d’une caméra par vision monoculaire. Dans la littérature, il existe différentes méthodes qui peuvent être classées en trois catégories. La première catégorie de méthodes considère une caméra évoluant dans un environnement complètement inconnu (SLAM). Cette méthode réalise une reconstruction enligne de primitives observées dans des images d’une séquence vidéo et utilise cette reconstruction pour localiser la caméra. Les deux autres permettent une localisation par rapport à un objet 3D de la scène en s’appuyant sur la connaissance, a priori, d’un modèle de cet objet (suivi basé modèle). L’une utilise uniquement l’information du modèle 3D de l’objet pour localiser la caméra, l’autre peut être considérée comme l’intermédiaire entre le SLAM et le suivi basé modèle. Cette dernière méthode consiste à localiser une caméra par rapport à un objet en utilisant, d’une part, le modèle de ce dernier et d’autre part, une reconstruction en ligne des primitives de l’objet d’intérêt. Cette reconstruction peut être assimilée à une mise à jour du modèle initial (suivi basé modèle avec mise à jour). Chacune de ces méthodes possède des avantages et des inconvénients. Dans le cadre de ces travaux de thèse, nous proposons une solution unifiant l’ensemble de ces méthodes de localisation dans un unique cadre désigné sous le terme de SLAM contraint. Cette solution, qui unifie ces différentes méthodes, permet de tirer profit de leurs avantages tout en limitant leurs inconvénients respectifs. En particulier, nous considérons que la caméra évolue dans un environnement partiellement connu, c’est-à-dire pour lequel un modèle (géométrique ou photométrique) 3D d’un objet statique de la scène est disponible. L’objectif est alors d’estimer de manière précise la pose de la caméra par rapport à cet objet 3D. L’information absolue issue du modèle 3D de l’objet d’intérêt est utilisée pour améliorer la localisation de type SLAM en incluant cette information additionnelle directement dans le processus d’ajustement de faisceaux. Afin de pouvoir gérer un large panel d’objets 3D et de scènes, plusieurs types de contraintes sont proposées dans ce mémoire. Ces différentes contraintes sont regroupées en deux approches. La première permet d’unifier les méthodes SLAM et de suivi basé modèle, en contraignant le déplacement de la caméra via la projection de primitives existantes extraites du modèle 3D dans les images. La seconde unifie les méthodes SLAM et de suivi basé modèle avec mise à jour en contraignant les primitives reconstruites par le SLAM à appartenir à la surface du modèle (unification SLAM et mise à jour du modèle). Les avantages de ces différents ajustements de faisceaux contraints, en terme de précision, de stabilité de recalage et de robustesse aux occultations, sont démontrés sur un grand nombre de données de synthèse et de données réelles. Des applications temps réel de réalité augmentée sont également présentées sur différents types d’objets 3D. Ces travaux ont fait l’objet de 4 publications internationales, de 2 publications nationales et d’un dépôt de brevet. / This thesis tackles the problem of real time location of a monocular camera. In the literature, there are different methods which can be classified into three categories. The first category considers a camera moving in a completely unknown environment (SLAM). This method performs an online reconstruction of the observed primitives in the images and uses this reconstruction to estimate the location of the camera. The two other categories of methods estimate the location of the camera with respect to a 3D object in the scene. The estimation is based on an a priori knowledge of a model of the object (Model-based). One of these two methods uses only the information of the 3D model of the object to locate the camera. The other method may be considered as an intermediary between the SLAM and Model-based approaches. It consists in locating the camera with respect to the object of interest by using, on one hand the 3D model of this object, and on the other hand an online reconstruction of the primitives of the latter. This last online reconstruction can be regarded as an update of the initial 3D model (Model-based with update). Each of these methods has advantages and disadvantages. In the context of this thesis, we propose a solution in order to unify all these localization methods in a single framework referred to as the constrained SLAM, by taking parts of their benefits and limiting their disadvantages. We, particularly, consider that the camera moves in a partially known environment, i.e. for which a 3D model (geometric or photometric) of a static object in the scene is available. The objective is then to accurately estimate the pose (position and orientation) of the camera with respect to this object. The absolute information provided by the 3D model of the object is used to improve the localization of the SLAM by directly including this additional information in the bundle adjustment process. In order to manage a wide range of 3D objets and scenes, various types of constraints are proposed in this study and grouped into two approaches. The first one allows to unify the SLAM and Model-based methods by constraining the trajectory of the camera through the projection, in the images, of the 3D primitives extracted from the model. The second one unifies the SLAM and Model-based with update methods, by constraining the reconstructed 3D primitives of the object to belong to the surface of the model (unification SLAM and model update). The benefits of the constrained bundle adjustment framework in terms of accuracy, stability, robustness to occlusions, are demonstrated on synthetic and real data. Real time applications of augmented reality are also presented on different types of 3D objects. This work has been the subject of four international publications, two national publications and one patent.

Un besoin applicatif existe en terme de localisation 3D d’objets par vision. Cette technologie devient en effet de plus en plus populaire dans le milieu industriel où elle peut être utile lors de contrôle qualité, de robotisation de tâches ou encore d’aide à la maintenance par Réalité Augmentée. Néanmoins, le déploiement de telles applications est actuellement limité en raison de la difficulté à allier qualité de localisation, facilité de mise en oeuvre et généricité de la solution. En effet, la majorité des solutions implique : soit des étapes de mise en oeuvre complexes comme avec l’installation de capteurs de mouvement ou une préparation supervisée du modèle CAO; soit un manque de précision de la localisation dans le cadre de certaines applications nécessitant de prendre en compte des mouvements de fortes amplitudes de la caméra (provoquant du flou de bouger et des tremblements dans le flux vidéo) ainsi que des occultations partielles ou totales de l’objet ; soit enfin une restriction sur la nature de l’objet, celui-ci devant être texturé, de petite taille ou encore polyédrique pour avoir une bonne localisation. La plupart des solutions de localisation existantes correspondent à des approches de suivi basé modèle. Cette méthode consiste à estimer la pose relative entre la caméra et l’objet d’intérêt par mises en correspondance de primitives 3D extraites du modèle avec des primitives 2D extraites d’images d’un flux vidéo. Pour autant, cette approche atteint ses limites lorsque l’objet est difficilement observable dans l’image.Afin d’améliorer la localisation lorsque l’application concerne un objet fixe, de récentes solutions se sont appuyées en complément des primitives du modèle, sur des primitives de l’environnement reconstruites au cours du processus de localisation. Ces approches combinent algorithmes de SLAM (Simultaneous Localization And Mapping) et de suivi d’objet basé contours en utilisant les informations du modèle comme contrainte dans le processus d’optimisation du SLAM. Pour cela, un terme d’erreur est ajouté à la fonction de coût classique.Celui-ci mesure l’erreur de re-projection entre des primitives 3D issues des arêtes franches du modèle et les points de contour 2D dans l’image qui leur sont associés. L’ajout de cette contrainte permet d’exprimer la localisation du SLAM dans le repère de l’objet d’intérêt tout en réduisant sa dérive. Les solutions de SLAM contraint au modèle n’exploitant cependant que les contours francs du modèle, ne sont pas génériques et ne permettent de localiser que des objets polyédriques. De plus, l’ajout de cette contrainte entraîne une forte augmentation de la consommation mémoire, les images de contours nécessaires à l’étape de mise en correspondance devant être conservées.Les travaux présentés dans ce mémoire de thèse visent à fournir une solution répondant simultanément à l’ensemble des besoins concernant la facilité de déploiement, la qualité de localisation et la généricité sur la nature des objets suivis. Aussi, notre solution basée sur un algorithme de SLAM visuel contraint basé images clés, se restreint-elle au seul usage d’une caméra couleur, les caméras RGBD impliquant généralement une limite sur le volume, la nature réflective ou absorbante de l’objet, et sur la luminosité de son environnement. Cette étude est en outre restreinte à la seule exploitation de modèles 3D géométrique non texturés, les textures pouvant difficilement être considérées comme stables dans le temps (usure, taches...) et pouvant varier pour un même objet manufacturé. De plus, les modèles à base de nuages de descripteurs locaux ou les modèles surfaciques texturés sont actuellement des données peu disponibles dans l’industrie. Enfin, nous faisons le choix d’estimer la pose de la caméra de manière géométrique et non par apprentissage. Le suivi d’objets à l’aide d’apprentissage automatique est en effet encore difficilement exploitable en milieu industriel. (...) / In the industry domain, applications such as quality control, automation of complex tasks or maintenance support with Augmented Reality (AR) could greatly benefit from visual tracking of 3D objects. However, this technology is under-exploited due to the difficulty of providing deployment easiness, localization quality and genericity simultaneously. Most existing solutions indeed involve a complex or an expensive deployment of motion capture sensors, or require human supervision to simplify the 3D model. And finally, most tracking solutions are restricted to textured or polyhedral objects to achieved an accurate camera pose estimation.Tracking any object is a challenging task due to the large variety of object forms and appearances. Industrial objects may indeed have sharp edges, or occluding contours that correspond to non-static and view-point dependent edges. They may also be textured or textureless. Moreover, some applications require to take large amplitude motions as well as object occlusions into account, tasks that are not always dealt with common model-based tracking methods. These approaches indeed exploit 3D features extracted from a model, that are matched with 2D features in the image of a video-stream. However the accuracy and robustness of the camera localization depend on the visibility of the object as well as on the motion of the camera. To better constrain the localization when the object is static, recent solutions rely on environment features that are reconstructed online, in addition to the model ones. These approaches combine SLAM (Simultaneous Localization And Mapping) and model-based tracking solutions by using constraints from the 3D model of the object of interest. Constraining SLAM algorithms with a 3D model results in a drift free localization. However, such approaches are not generic since they are only adapted for textured or polyhedral objects. Furthermore, using the 3D model to constrain the optimization process may generate high memory consumption,and limit the optimization to a temporal window of few cameras. In this thesis, we propose a solution that fulfills the requirements concerning deployment easiness, localization quality and genericity. This solution, based on a visual key-frame-based constrained SLAM, only exploits an RGB camera and a geometric CAD model of the static object of interest. An RGB camera is indeed preferred over an RGBD sensor, since the latter imposes limits on the volume, the reflectiveness or the absorptiveness of the object, and the lighting conditions. A geometric CAD model is also preferred over a textured model since textures may hardly be considered as stable in time (deterioration, marks,...) and may vary for one manufactured object. Furthermore, textured CAD models are currently not widely spread. Contrarily to previous methods, the presented approach deals with polyhedral and curved objects by extracting dynamically 3D contour points from a model rendered on GPU. This extraction is integrated as a structure constraint into the constrained bundle adjustment of a SLAM algorithm. Moreover we propose different formalisms of this constraint to reduce the memory consumption of the optimization process. These formalisms correspond to hybrid structure/trajectory constraints, that uses output camera poses of a model-based tracker. These formalisms take into account the structure information given by the 3D model while relying on the formalism of trajectory constraints. The proposed solution is real-time, accurate and robust to occlusion or sudden motion. It has been evaluated on synthetic and real sequences of different kind of objects. The results show that the accuracy achieved on the camera trajectory is sufficient to ensure a solution perfectly adapted for high-quality Augmented Reality experiences for the industry.