View-based slam in the 2-1/2d space with omnidirectional images and feature point information

Resumen

Mobile robotics has experienced an important proliferation in the recent days, with many fields of application available. A great variety of different mobile robots are present in different sectors of society, most of which are designated as autonomous. This term implies that the robot manages to operate itself, without any special super- vision. To that purpose, the robot must be enabled to gather information from the environment in order to build its own understanding, which yields a map estimation. The scope of this thesis is focused on this aspect: the map building process with visual information from the environment. This process entails a non-trivial task, since it poses a challenge when it comes to obtaining a simultaneous estimation of the localization of the robot, and also of the map. This leads to one of the most essential paradigms in such context: the problem of SLAM (Simultaneous Localization And Mapping). Different sort of information can be acquired by a set of well known sensors boarded on the robot, such as laser, sonar, GPS, etc. However, digital cameras have arisen as a promising alternative. They provide low consumption, low cost and lightness. More- over, these visual sensors represent a potential tool for encoding large amounts of information within an only image. Thus in this work we propose a new map model, embedded in a visual SLAM approach, which is solely based on the use of omnidirectional images acquired with a monocular camera. An important strength of this camera resides on its particular wide field of view. In addition, we process the information extracted from feature points, as physical landmarks which are visually detected on the images. This idea differs from traditional approaches, which basically concentrate on the accumulative scheme for the incremental re-estimation of all the landmarks in the map. Regarding the core algorithms under this context of visual SLAM, this thesis pro- poses several improvements to the robustness of the standard algorithm models. In particular, we present a customized offline model, which is capable of reducing those harmful effects associated with non-linear noise, as those introduced by catadioptric cameras. Many of the most accepted approaches are highly sensitive to this effects and fail to provide convergence assurance for the final estimation. Moreover, another recognized drawback of former approaches is the management of the uncertainty of the system. This is usually originated by the same non-linear sources. Consequently, the estimation may be severely impaired as errors dramatically compromise its convergence. In this sense, this thesis contributes to the achievement of a robust model for uncertainty reduction, which is dynamically devised. As a general commitment along all this thesis, we establish an experimental frame- work for all the different approaches and contributions made as a result of the research conducted in this context. Thus both simulated and real dataset experiments are repeatedly presented along this document.