Visual recognition for localization purposes using omnidirectional images maps

Abstract

This thesis is an exploratory work with the objective of developing techniques for recognition using visual maps constituted by paracatadioptric images. This kind of images contain a more complete description of the scene, covering a wider area than e.g. perspective images. However, the radial distortion they present is high which difficult their usage. The main challenge of the thesis is establishing correspondences between a query image captured by a standard camera and the paracatadioptric images. Different types of rectification strategies are studied in order to correct the radial distortion present in the paracatadioptric images. Matching between perspective images and rectified images from the paracatadioptric images is performed, using SIFT algorithm [18]. Also, a simple procedure to roughly estimate the calibration matrix of the system is explained. Additionally, several modifications to the original SIFT algorithm are proposed, including a change in the initial scale for the Gaussian blurring and a novel approach for feature detection and description of stable local features, directly over the paracatadioptric images (based on [17]). With this approach, the detection is carried in a scale-space image representation built using an adaptive Gaussian filter that takes into account the geometry of the paracatadioptric system. As an additional result, a brief study of feature detection and matching between paracatadioptric images is performed. A new mapping function for the adaptive filtering is tested, outperforming the other approaches. Another important topic of the thesis is to perform image based localization using a database of omnidirectional images. A recognition scheme using an hierarchical vocabulary tree is built, based on [25]. Different visual vocabularies and training data are used. Additionally, several methods for feature extraction in the database images are analyzed, including the original SIFT algorithm and SIFT with implicit filtering. Finally, methods to improve the retrieval performance are studied. To encode more spatial information in the searching step, the concept of geometry-preserving visual phrases is used [33]. Additionally, to provide a more precise ranking of the retrieved images, a geometric consistency check (using RANSAC) is performed on the top-ranked images.