A large number of image registration techniques have been developed for various types of sensors and applications, with the aim to improve the accuracy, computational complexity, generality, and robustness. They can be broadly classified into two categories: intensity-based and feature-based methods. The primary drawback of the intensity-based approaches is that it may fail unless the two images are misaligned by a moderate difference in scale, rotation, and translation. In addition, intensity-based methods lack the robustness in the presence of non-spatial distortions due to different imaging conditions between images. In this dissertation, the image registration is formulated as a two-stage hybrid approach combining both an initial matching and a final matching in a coarse-to-fine manner. In the proposed hybrid framework, the initial matching algorithm is applied at the coarsest scale of images, where the approximate transformation parameters could be first estimated. Subsequently, the robust gradient-based estimation algorithm is incorporated into the proposed hybrid approach using a multi-resolution scheme. Several novel and effective initial matching algorithms have been proposed for the first stage. The variations of the intensity characteristics between images may be large and non-uniform because of non-spatial distortions. Therefore, in order to effectively incorporate the gradient-based robust estimation into our proposed framework, one of the fundamental questions should be addressed: what is a good image representation to work with using gradient-based robust estimation under non-spatial distortions. With the initial matching algorithms applied at the highest level of decomposition, the proposed hybrid approach exhibits superior range of convergence. The gradient-based algorithms in the second stage yield a robust solution that precisely registers images with sub-pixel accuracy. A hierarchical iterative searching further enhances the convergence range and rate. The simulation results demonstrated that the proposed techniques provide significant benefits to the performance of image registration.

“Smart” surveillance systems require a visual tracking system that is able to detect and follow a moving target in the field of view of a camera. Visual tracking systems have been traditionally developed either as application specific hardware or as software written for parallel architectures because of the large number of computations that have to be performed at very high speeds. This thesis describes the implementations of two visual tracking systems on a custom computing machine based on Field Programmable Gate Arrays (FPGAs). The implementations apply a coarse-to-fine search on Gaussian pyramids constructed from the images generated by a camera. One system tracks a target of size 16x16 in an image sequence with output images of size 256x256. This system is capable of operating at 30 pyramids per second. The second system tracks a target of size 16x16 in an image sequence with output images of size 512x512. This system is capable of operating at 15 pyramids per second. Both systems are designed with pipelined architectures and numerical computations are handled using a SIMD approach. / Master of Science

L'objectif de cette thèse est l'étude d'un descripteur d'images adapté à une grande variété d'applications. Nous cherchons à obtenir un descripteur robuste et discriminant, facile à adapter et peu coûteux en calcul et en mémoire.Nous définissons un nouveau descripteur, composé de valeurs du Laplacien à différentes échelles et de valeurs d'une transformée de Fourier radiale, calculées à partir d'une pyramide Gaussienne. Ce descripteur capture une information de forme multi-échelle autour d'un point de l'image. L'expérimentation a montré que malgré une taille mémoire réduite les performances en robustesse et en pouvoir discriminant de ce descripteur sont à la heuteur de l'état de l'art.Nous avons expérimenté ce descripteur avec trois types de tâches différentes.Le premier type de tâche est la mise en correspondance de points-clés avec des images transformées par rotation, changement d'échelle, floutage, codage JPEG, changement de point de vue, ou changement d'éclairage. Nous montrons que la performance de notre descripteur est au niveau des meilleurs descripteurs connus dans l'état de l'art. Le deuxième type de tâche est la détection de formes. Nous avons utilisé le descripteur pour la création de deux détecteurs de personnes, construits avec Adaboost. Comparé à un détecteur semblable construit avec des histogrammes de gradients (HOG) nos détecteurs sont très compétitifs tout en utilisant des descripteurs sensiblement plus compacts. Le dernier type de tâche est la détection de symétries de réflexion dans des images "du monde réel". Nous proposons une technique de détection d'axes potentiels de symétries en miroir. Avec cette tâche nous montrons que notre descripteur peut être genéralisé à des situations complexes. L'expérimentation montre que cette méthode est robuste et discriminante, tout en conservant un faible coût en calcul et en mémoire. / In this thesis we explore a new image description method composed of a multi-scale vector of Laplacians of Gaussians, the Laplacian Profile, and a Radial Fourier Transform. This method captures shape information with different proportions around a point in the image. A Gaussian pyramid of scaled images is used for the extraction of the descriptor vectors. The aim of this new method is to provide image description that can be suitable for diverse applications. Adjustability as well as low computational and memory needs are as important as robustness and discrimination power. We created a method with the ability to capture the image signal efficiently with descriptor vectors of particularly small length compared to the state of the art. Experiments show that despite its small vector length, the new descriptor shows reasonable robustness and discrimination power that are competitive to the state of the art performance.We test our proposed image description method on three different visual tasks. The first task is keypoint matching for images that have undergone image transformations like rotation, scaling, blurring, JPEG compression, changes in viewpoint and changes in light. We show that against other methods from the state of the art, the proposed descriptor performs equivalently with a very small vector length. The second task is on pattern detection. We use the proposed descriptor to create two different Adaboost based detectors for people detection in images. Compared to a similar detector using Histograms of Oriented Gradients (HOG), the detectors with the proposed method show competitive performance using significantly smaller descriptor vectors. The last task is on reflection symmetry detection in real world images. We introduce a technique that exploits the proposed descriptor for detecting possible symmetry axes for the two reflecting parts of a mirror symmetric pattern. This technique introduces constraints and ideas of how to collect more efficiently the information that is important to identify reflection symmetry in images. With this task we show that the proposed descriptor can be generalized for rather complicated applications. The set of the experiments confirms the qualities of the proposed method of being easily adjustable and requires relatively low computational and storage requirements while remaining robust and discriminative.