This work focuses on analysing and improving feature detection and matching. After creating an initial framework of study, four main areas of work are researched. These areas make up the main chapters within this thesis and focus on using the Scale Invariant Feature Transform (SIFT).The preliminary analysis of the SIFT investigates how this algorithm functions. Included is an analysis of the SIFT feature descriptor space and an investigation into the noise properties of the SIFT. It introduces a novel use of the a contrario methodology and shows the success of this method as a way of discriminating between images which are likely to contain corresponding regions from images which do not. Parameter analysis of the SIFT uses both parameter sweeps and genetic algorithms as an intelligent means of setting the SIFT parameters for different image types utilising a GPGPU implementation of SIFT. The results have demonstrated which parameters are more important when optimising the algorithm and the areas within the parameter space to focus on when tuning the values. A multi-exposure, High Dynamic Range (HDR), fusion features process has been developed where the SIFT image features are matched within high contrast scenes. Bracketed exposure images are analysed and features are extracted and combined from different images to create a set of features which describe a larger dynamic range. They are shown to reduce the effects of noise and artefacts that are introduced when extracting features from HDR images directly and have a superior image matching performance. The final area is the development of a novel, 3D-based, SIFT weighting technique which utilises the 3D data from a pair of stereo images to cluster and class matched SIFT features. Weightings are applied to the matches based on the 3D properties of the features and how they cluster in order to attempt to discriminate between correct and incorrect matches using the a contrario methodology. The results show that the technique provides a method for discriminating between correct and incorrect matches and that the a contrario methodology has potential for future investigation as a method for correct feature match prediction.

Ces travaux s’inscrivent dans le cadre du projet VIPA « Véhicule Individuel Public Autonome », au cours duquel le LASMEA et ses partenaires ont mis au point des véhicules capables de naviguer automatiquement, sans aucune infrastructure extérieure dédiée, dans des zones urbaines (parkings, zones piétonnes, aéroports). Il est doté de deux caméras, l’une à l’avant, et l’autre à l’arrière. Avant son déploiement, le véhicule doit tout d’abord être étalonné et conduit manuellement afin de reconstruire la carte d’amers visuels dans laquelle il naviguera ensuite automatiquement. Les travaux de cette thèse ont pour but de développer et de mettre en oeuvre des méthodes souples permettant d’étalonner cet ensemble de caméras dont les champs de vue sont totalement disjoints. Après une étape préalable d’étalonnage intrinsèque et un état de l’art sur les systèmes multi-caméra, nous développons et mettons en oeuvre différentes méthodes d’étalonnage extrinsèque (déterminant les poses relatives des caméras à champs de vue disjoints). La première méthode présentée utilise un miroir plan pour créer un champ de vision commun aux différentes caméras. La seconde approche consiste à manoeuvrer le véhicule pendant que chaque caméra observe une scène statique composée de cibles (dont la détection est sous-pixellique). Dans la troisième approche, nous montrons que l’étalonnage extrinsèque peut être obtenu simultanément à la reconstruction 3D (par exemple lors de la phase d’apprentissage), en utilisant des points d’intérêt comme amers visuels. Pour cela un algorithme d’ajustement de faisceaux multi-caméra a été développé avec une implémentation creuse. Enfin, nous terminons par un étalonnage déterminant l’orientation du système multi-caméra par rapport au véhicule. / My research was involved in the VIPA « Automatic Electric Vehicle for Passenger Transportation » project. During which, the LASMEA and its partnerships have developed vehicles able to navigate autonomously, without any outside dedicated infrastructure in an urban environment (parking lots, pedestrian areas, airports). Two cameras are rigidly embedded on a vehicle : one at the front, another at the back. Before being available for autonomous navigation tasks, the vehicle have to be calibrated and driven manually in order to build a visual 3D map (calibration and learning steps). Then, the vehicle will use this map to localize itself and drive autonomously. The goals of this thesis are to develop and apply user friendly methods, which calibrate this set of nonoverlapping cameras. After a first step of intrinsic calibration and a state of the art on multi-camera rigs, we develop and test several methods to extrinsically calibrate non-overlapping cameras (i.e. estimate the camera relative poses). The first method uses a planar mirror to create an overlap between views of the different cameras. The second procedure consists in manoeuvring the vehicle while each camera observes a static scene (composed of a set of targets, which are detected accurately). In a third procedure, we solve the 3D reconstruction and the extrinsic calibration problems simultaneously (the learning step can be used for that purpose) relying on visual features such as interest points. To achieve this goal a multi-camera bundle adjustment is proposed and implemented with a sparse data structures. Lastly, we present a calibration of the orientation of a multi-camera rig relative to the vehicle.

This diploma thesis deals with 3D reconstruction of objects using image analysis methods. The work includes mathematical theory associated with this problem, a procedure for creating 2D sharp images and 3D reconstruction itself. The outputs are 2D sharp images, 3D models, stl models. Different kinds of data are analyzed.

V posledním desetiletí došlo ke zvýšení zájmu o užití 3D dat k biometrické identifikaci osob. Možná vůbec největší výzkum proběhl v oblasti 3D rozpoznávání podle obličeje, přičemž je v současné době dostupných vícero komerčních zařízení. V oblastni rozpoznávání podle 3D geometrie ruky byl v minulých letech proveden určitý výzkum jehož výsledkem však nebylo žádné komerční zařízení. Nezávisle na tomto výzkumu se v posledních letech velmi rozšířil trh s cenově dostupnými 3D sensory, což potenciálně umožňuje jejich nasazení v mnoha typech biometrických systémů. Hlavním cílem této práce je vytvořit funkční vzorek bezdotykového systému pro rozpoznávání osob podle 3D geometrie ruky, který bude používat novou levnou kameru RealSense 3D vyvíjenou v současné době firmou Intel. Jedním z problémů při použití RealSense kamery je její velmi malý form factor, který je příčinou nižší kvality výsledných snímků v porovnání s velmi drahými alternativami, které byly použity v již dříve zmíněném výzkumu 3D biometrických systémů. Práce se snaží analyzovat robustnost různých 2D a 3D příznaků a vyzkoušet několik různých přístupů k jejich fúzi. Rovněž je vyhodnocena výkonnost výsledného systému, kde je ukázáno, že navržené řešení dosahuje výsledků porovnatelných se state-of-the-art.

2D to 3D reconstruction is an ill-posed problem in the field of Autonomous Robot Navigation. Many practitioners are tend to utilize the enormous success of Deep Learning techniques like CNN, ANN etc to solve tasks related to this 2D to 3D reconstruction. Generally, every deep learning model involves implementation of different optimizers related to the tasks to lower the possible negativity in its results and selection of hyper parameter values for these optimizers during the process of training the model with required dataset.Selection of this optimizer hyper-parameters requires in-depth knowledge and trials and errors. So proposing optimal hyper parameters for optimizers results in no waste in computational resources and time.Hence solution for the selected task cab found easily. The main objective of this research is to propose optimal hyper parameter values of various deep learning optimizers related to 2D to 3D reconstruction and proposing best optimizer among them in terms of computational time and resources To achieve the goal of this study two research methods are used in our work. The first one is a Systematic Literature Review; whose main goal is to reveal the widely selected and used optimizers for 2D to 3D reconstruction model using 3D Deep Learning techniques.The second, an experimental methodology is deployed, whose main goal is to propose the optimal hyper parameter values for respective optimizers like Adam, SGD+Momentum, Adagrad, Adadelta and Adamax which are used in 3D reconstruction models. In case of the computational time, Adamax optimizer outperformed all other optimizers used with training time (1970min), testing time (3360 min), evaluation-1 (16 min) and evaluation-2 (14 min).In case of Average Point cloud points, Adamax outperformed all other optimizers used with Mean value of 28451.04.In case of pred->GT and GT->pred values , Adamax optimizer outperformed all other optimizers with mean values of 4.742 and 4.600 respectively. Point Cloud Images with respective dense cloud points are obtained as results of our experiment.From the above results,Adamax optimizer is proved to be best in terms of visualization of Point Cloud images with optimal hyper parameter values as below:Epochs : 1000    Learning Rate : 1e-2    Chunk size : 32    Batch size : 32.  In this study,'Adamax' optimizer with optimal hyper para meter values and better Point Cloud Image is proven to be the best optimizer that can be used in a 2D to 3D reconstruction related task that deals with Point Cloud images

Dissertation deals with analysis and design of optoelectronic and photogrammetric measuring systems. Specific design of optoelectronic contactless flat object area meters with analysis of attainable measurement accuracy is described. Next part is dedicated to stereophotogrammetry - principles of 3D reconstructions, methods of camera self-calibration and matching points in images are described. The analysis of attainable accuracy of monitored parameters is discussed too. Finally, the test program with implemented described routines is introduced. This test program enables practical aplication of stereophotogrammetric system for taking spatial coordinates of 3D objects.

This work describes 3D reconstruction using stereo vision. It presents methods for automatic localization of corresponding points in both images and their reprojection into 3D space. Application created can be used for navigation of a robot and object avoidance. Second part of the document describes chosen components of the robot. Path finding algorithms are also discussed, particulary Voronoi's diagram.

This thesis focuses on reconstruction of 3D environment model from a set of photographs followed by camera localization. It describes basic principles and techniques used to create environmental models and techniques for camera pose estimation from 2D camera points to 3D model points. It also examines the influence of parameters on the quality of reconstruction and the possibilities of localization. It compares the quality of the descriptors in the process of creation of the model and based on localization it allows to implement augmented reality.

Existing systems of scene reconstruction and theorethical basics necessary for scene reconstruction from images data are described in this work. System of scene reconstruction from video was designed and implemented. Its results were analyzed and possible future work was proposed. OpenCV, ArtToolKit and SIFT libraries which were used in this project are also described.

Dans un contexte industriel, la tomographie 3D par rayons X vise à imager virtuellement une pièce afin d'en contrôler l'intérieur. Le volume virtuel de la pièce est obtenu par un algorithme de reconstruction, prenant en entrées les projections de rayons X qui ont été envoyés à travers la pièce. Beaucoup d'incertitudes résident dans ces projections à cause de phénomènes non contrôlés tels que la diffusion et le durcissement de faisceau, causes d'artefacts dans les reconstructions conventionnelles par rétroprojection filtrée. Afin de compenser ces incertitudes, les méthodes de reconstruction dites itératives tentent de faire correspondre la reconstruction à un modèle a priori, ce qui, combiné à l'information apportée par les projections, permet d'améliorer la qualité de reconstruction. Dans ce contexte, cette thèse propose de nouvelles méthodes de reconstruction itératives pour le contrôle de pièces produites par le groupe SAFRAN. Compte tenu de nombreuses opérations de projection et de rétroprojection modélisant le processus d'acquisition, les méthodes de reconstruction itératives peuvent être accélérées grâce au calcul parallèle haute performance sur processeur graphique (GPU). Dans cette thèse, les implémentations sur GPU de plusieurs paires de projecteur-rétroprojecteur sont décrites. En particulier, une nouvelle implémentation pour la paire duale dite à empreinte séparable est proposée. Beaucoup de pièces produites par SAFRAN pouvant être vues comme des volumes constants par morceaux, un modèle a priori de Gauss-Markov-Potts est introduit, à partir duquel est déduit un algorithme de reconstruction et de segmentation conjointes. Cet algorithme repose sur une approche bayésienne permettant d'expliquer le rôle de chacun des paramètres. Le caractère polychromatique des rayons X par lequel s'expliquent la diffusion et le durcissement de faisceau est pris en compte par l'introduction d'un modèle direct séparant les incertitudes sur les projections. Allié à un modèle de Gauss-Markov-Potts sur le volume, il est montré expérimentalement que ce nouveau modèle direct apporte un gain en précision et en robustesse. Enfin, l'estimation des incertitudes sur la reconstruction est traitée via l'approche bayésienne variationnelle. Pour obtenir cette estimation en un temps de calcul raisonnable, il est montré qu'il est nécessaire d'utiliser une paire duale de projecteur-rétroprojecteur. / In industry, 3D X-ray Computed Tomography aims at virtually imaging a volume in order to inspect its interior. The virtual volume is obtained thanks to a reconstruction algorithm based on projections of X-rays sent through the industrial part to inspect. In order to compensate uncertainties in the projections such as scattering or beam-hardening, which are cause of many artifacts in conventional filtered backprojection methods, iterative reconstruction methods bring further information by enforcing a prior model on the volume to reconstruct, and actually enhance the reconstruction quality. In this context, this thesis proposes new iterative reconstruction methods for the inspection of aeronautical parts made by SAFRAN group. In order to alleviate the computational cost due to repeated projection and backprojection operations which model the acquisition process, iterative reconstruction methods can take benefit from the use of high-parallel computing on Graphical Processor Unit (GPU). In this thesis, the implementation on GPU of several pairs of projector and backprojector is detailed. In particular, a new GPU implementation of the matched Separable Footprint pair is proposed. Since many of SAFRAN's industrial parts are piecewise-constant volumes, a Gauss-Markov-Potts prior model is introduced, from which a joint reconstruction and segmentation algorithm is derived. This algorithm is based on a Bayesian approach which enables to explain the role of each parameter. The actual polychromacy of X-rays, which is responsible for scattering and beam-hardening, is taken into account by proposing an error-splitting forward model. Combined with Gauss-Markov-Potts prior on the volume, this new forward model is experimentally shown to bring more accuracy and robustness. At last, the estimation of the uncertainties on the reconstruction is investigated by variational Bayesian approach. In order to have a reasonable computation time, it is highlighted that the use of a matched pair of projector and backprojector is necessary.