We present a statistical image-based shape + structure model for Bayesian visual hull reconstruction and 3D structure inference. The 3D shape of a class of objects is represented by sets of contours from silhouette views simultaneously observed from multiple calibrated cameras. Bayesian reconstructions of new shapes are then estimated using a prior density constructed with a mixture model and probabilistic principal components analysis. We show how the use of a class-specific prior in a visual hull reconstruction can reduce the effect of segmentation errors from the silhouette extraction process. The proposed method is applied to a data set of pedestrian images, and improvements in the approximate 3D models under various noise conditions are shown. We further augment the shape model to incorporate structural features of interest; unknown structural parameters for a novel set of contours are then inferred via the Bayesian reconstruction process. Model matching and parameter inference are done entirely in the image domain and require no explicit 3D construction. Our shape model enables accurate estimation of structure despite segmentation errors or missing views in the input silhouettes, and works even with only a single input view. Using a data set of thousands of pedestrian images generated from a synthetic model, we can accurately infer the 3D locations of 19 joints on the body based on observed silhouette contours from real images.

Nous nous intéressons à l'acquisition automatique de mouvements 3D de personnes. Cette opération doit être réalisée sans un équipement spécialisé (marqueurs ou habillage spécifique), pour rendre son utilisation générale, sous la contrainte du temps réel. Pour répondre à ces questions, nous sommes amenés à traiter de la reconstruction et l'analyse de la forme 3D. Concernant le problème de reconstruction 3D en temps réel d'entités en mouvement à partir de plusieurs vues, les approches existantes font souvent appel à des calculs complexes incompatibles avec la contrainte du temps réel. Les approches du type SFS offrent un compromis intéressant entre efficacité algorithmique et précision. Ces dernières utilisent les silhouettes issues de chaque caméra pour proposer un volume englobant des objets. Cependant elles nécessitent un environnement particulièrement contraint, dont le placement minutieux des caméras. Les travaux présentés dans ce manuscrit généralisent l'utilisation des approches SFS à des environnements peu contrôlés. L'acquisition du mouvement revient à déterminer les paramètres offrant la meilleure corrélation entre le modèle et la reconstruction 3D. Notre objectif étant le suivi temps réel, nous proposons des méthodes qui offrent la précision requise et le temps réel. Couplé à un suivi temporel par filtre de Kalman, à un recalage d'objets géométriques simples (ellipsoïdes, sphères, etc.), nous proposons un système temps réel, offrant une erreur de l'ordre de 6%.De par sa robustesse, il permet le suivi simultané de plusieurs personnes, même lors de contacts. Les résultats obtenus ouvrent des perspectives à un transfert vers des applications grand public / We aim at automatically capturing 3D motion of persons without markers. To make it flexible, and to consider interactive applications, we address real-time solution, without specialized instrumentation. Real-time body estimation and shape analyze lead to home motion capture application. We begin by addressing the problem of 3D real-time reconstruction of moving objects from multiple views. Existing approaches often involve complex computation methods, making them incompatible with real-time constraints. Shape-From-Silhouette (SFS) approaches provide interesting compromise between algorithm efficiency and accuracy. They estimate 3D objects from their silhouettes in each camera. However they require constrained environments and cameras placement. The works presented in this document generalize the use of SFS approaches to uncontrolled environments. The main methods of marker-less motion capture, are based on parametric modeling of the human body. The acquisition of movement goal is to determine the parameters that provide the best correlation between the model and the 3D reconstruction.The following approaches, more robust, use natural markings of the body extremities: the skin. Coupled with a temporal Kalman filter, a registration of simple geometric objects, or an ellipsoids' decomposition, we have proposed two real-time approaches, providing a mean error of 6%. Thanks to the approach robustness, it allows the simultaneous monitoring of several people even in contacts. The results obtained open up prospects for a transfer to home applications

<p>This thesis presents a complete and fully automatic method for estimating the volume of an airbag, through all stages of its inflation, with multiple synchronized high-speed cameras.</p><p>Using recorded contours of the inflating airbag, its visual hull is reconstructed with a novel method: The intersections of all back-projected contours are first identified with an accelerated epipolar algorithm. These intersections, together with additional points sampled from concave surface regions of the visual hull, are then Delaunay triangulated to a connected set of tetrahedra. Finally, the visual hull is extracted by carving away the tetrahedra that are classified as inconsistent with the contours, according to a voting procedure.</p><p>The volume of an airbag's visual hull is always larger than the airbag's real volume. By projecting a known synthetic model of the airbag into the cameras, this volume offset is computed, and an accurate estimate of the real airbag volume is extracted. </p><p>Even though volume estimates can be computed for all camera setups, the cameras should be specially posed to achieve optimal results. Such poses are uniquely found for different airbag models with a separate, fully automatic, simulated annealing algorithm.</p><p>Satisfying results are presented for both synthetic and real-world data.</p>

This thesis presents a complete and fully automatic method for estimating the volume of an airbag, through all stages of its inflation, with multiple synchronized high-speed cameras. Using recorded contours of the inflating airbag, its visual hull is reconstructed with a novel method: The intersections of all back-projected contours are first identified with an accelerated epipolar algorithm. These intersections, together with additional points sampled from concave surface regions of the visual hull, are then Delaunay triangulated to a connected set of tetrahedra. Finally, the visual hull is extracted by carving away the tetrahedra that are classified as inconsistent with the contours, according to a voting procedure. The volume of an airbag's visual hull is always larger than the airbag's real volume. By projecting a known synthetic model of the airbag into the cameras, this volume offset is computed, and an accurate estimate of the real airbag volume is extracted. Even though volume estimates can be computed for all camera setups, the cameras should be specially posed to achieve optimal results. Such poses are uniquely found for different airbag models with a separate, fully automatic, simulated annealing algorithm. Satisfying results are presented for both synthetic and real-world data.

La reconstruction précise d’une scène 3D à partir de plusieurs caméras offre un contenu synthétique 3D à destination de nombreuses applications telles que le divertissement, la télévision et la production cinématographique. Cette thèse propose une nouvelle approche pour la reconstruction 3D multi-vues basée sur l’enveloppe visuelle et la stéréovision multi-oculaire. Cette approche nécessite en entrée l’enveloppe visuelle et plusieurs jeux d’images rectifiées issues de différentes unités multiscopiques constituées chacune de plusieurs caméras alignées et équidistantes. Nos contributions se situent à différents niveaux. Le premier est notre méthode de stéréovision multi-oculaire qui est fondée sur un nouvel échantillonnage de l’espace scénique et fournit une carte de matérialité exprimant la probabilité pour chaque point d’échantillonnage 3D d’appartenir à la surface visible par l’unité multiscopique. Le second est l’hybridation de cette méthode avec les informations issues de l’enveloppe visuelle et le troisième est la chaîne de reconstruction basée sur la fusion des différentes enveloppes creusées tout en gérant les informations contradictoires qui peuvent exister. Les résultats confirment : i) l’efficacité de l’utilisation de la carte de matérialité pour traiter les problèmes qui se produisent souvent dans la stéréovision, en particulier pour les régions partiellementoccultées ; ii) l’avantage de la fusion des méthodes de l’enveloppe visuelle et de la stéréovision multi-oculaire pour générer un modèle 3D précis de la scène. / Accurate reconstruction of a 3D scene from multiple cameras offers 3D synthetic content tobe used in many applications such as entertainment, TV, and cinema production. This thesisis placed in the context of the RECOVER3D collaborative project, which aims is to provideefficient and quality innovative solutions to 3D acquisition of actors. The RECOVER3Dacquisition system is composed of several tens of synchronized cameras scattered aroundthe observed scene within a chromakey studio in order to build the visual hull, with severalgroups laid as multiscopic units dedicated to multi-baseline stereovision. A multiscopic unitis defined as a set of aligned and evenly distributed cameras. This thesis proposes a novelframework for multi-view 3D reconstruction relying on both multi-baseline stereovision andvisual hull. This method’s inputs are a visual hull and several sets of multi-baseline views.For each such view set, a multi-baseline stereovision method yields a surface which is usedto carve the visual hull. Carved visual hulls from different view sets are then fused iterativelyto deliver the intended 3D model. Furthermore, we propose a framework for multi-baselinestereo-vision which provides upon the Disparity Space (DS), a materiality map expressingthe probability for 3D sample points to lie on a visible surface. The results confirm i) theefficient of using the materiality map to deal with commonly occurring problems in multibaselinestereovision in particular for semi or partially occluded regions, ii) the benefit ofmerging visual hull and multi-baseline stereovision methods to produce 3D objects modelswith high precision.