Contribuciones a la estimación de la pose de cámara

Résumé

Camera pose estimation is the problem of finding the orientation and localization of a camera with respect to an arbitrary coordinate system. Image-based solutions for this problem are an interesting option because its reduced cost. However, their main drawback is that the accuracy of the results is afected by the presence of noise in the images. The use of images for the camera pose estimation task is strongly related to the Perspective-n-Point (PnP) and Bundle Adjustment problem. Given a set of n correspondences between 3D points and its 2D projections on the image, PnP methods provide estimations of the camera pose. In addition, when the information about the 3D positions is unknow but a set of 2D projections taken from diferent viewpoints of the same 3D point are known, Bundle Adjustment methods are capable of finding simultaneously the 3D position of the points and the camera pose. Then the task of finding correspondences between 3D points and its 2D projections, and between 2D projections of diferent images is a fundamental step for the above mentioned problems. This PhD Thesis proposes two novel approaches to solve the problem of finding correspondeces using both natural and artificial features. In our first contribution, based on natural features, we propose a novel approach to find 2D correspondeces between images by a novel fusion approach combining information provided by several descriptors using the Dempster-Shafer Theory. The proposed method is able to fuse diferent sources of information considering their relative confidence in order to provide a better solution. Our second contribution focuses on the problem of nding the 2D projections of 3D points. We propose a novel approach for identification of artificial landmarks, which are a very popular method when robustness and speed are required. In particular, we propose to tackle the marker identi cation problem as a classi cation one. As a consequence, we develop methods able to detect such markers in complex real situations such as blurring and non-uniform lightning. The two contributions made in this Thesis have been compared with the state-of-art methods showing statistically significant improvements.