For robotic systems to realise lifelong autonomy they must be able to navigate accurately in changing environments. In this thesis we describe, implement and validate a new approach to the problem of long-term navigation. To begin, we present our stereo visual odometry system which provides highly accurate pose estimation. Our approach combines several techniques found in existing implementations and a recently published image descriptor that simplifies the solution architecture. The performance and versatility of our system is demonstrated through testing on multiple datasets. Equipped with our visual odometry system, we describe a new approach to the problem of lifelong navigation. We learn a model whose complexity varies naturally in accordance with the variation of scene appearance. As the robot repeatedly traverses its workspace, it accumulates distinct visual experiences that, in concert, implicitly represent the scene variation - each experience captures a visual mode. When operating in a previously visited area, we continually try to localise in these previous experiences while simultaneously running the visual odometry. Failure to localise in a sufficient number of prior experiences indicates an insufficient model of the workspace and instigates the laying down of the live image sequence as a new distinct experience. In this way, over time we can capture the typical temporally varying appearance of an environment and the number of experiences required tends to a constant. Although we focus on vision as a primary sensor, the ideas we present here are equally applicable to other sensor modalities. We demonstrate our approach working on a road vehicle operating over a three month period at different times of day, in different weather and lighting conditions.

Thesis (M.Eng.). / Title from title page of PDF (viewed 2008/01/30). Written for the Dept. of Biomedical Engineering. Includes bibliographical references.

Made available in DSpace on 2014-12-17T14:55:40Z (GMT). No. of bitstreams: 1 RafaelBG.pdf: 2393943 bytes, checksum: 45924e13c3c73c1eaaf09dcc478bd70e (MD5) Previous issue date: 2009-08-07 / Conselho Nacional de Desenvolvimento Cient?fico e Tecnol?gico / We propose a new approach to reduction and abstraction of visual information for robotics vision applications. Basically, we propose to use a multi-resolution representation in combination with a moving fovea for reducing the amount of information from an image. We introduce the mathematical formalization of the moving fovea approach and mapping functions that help to use this model. Two indexes (resolution and cost) are proposed that can be useful to choose the proposed model variables. With this new theoretical approach, it is possible to apply several filters, to calculate disparity and to obtain motion analysis in real time (less than 33ms to process an image pair at a notebook AMD Turion Dual Core 2GHz). As the main result, most of time, the moving fovea allows the robot not to perform physical motion of its robotics devices to keep a possible region of interest visible in both images. We validate the proposed model with experimental results / N?s propomos uma nova abordagem para reduzir e abstrair informa??es visuais para aplica??es de vis?o rob?tica. Basicamente, usamos uma representa??o emmulti-resolu??o em combina??o com uma f?vea m?vel para reduzir a quantidade de informa??es de uma imagem. Apresentamos a formaliza??o matem?tica do modelo em conjunto com fun??es de mapeamento que auxiliam na utiliza??o do modelo. Propomos dois ?ndices (resolu??o e custo) que visam auxiliar na escolha das vari?veis do modelo proposto. Com essa nova abordagem te?rica, ? poss?vel aplicar diversos filtros, calcular disparidade est?reo e obter an?lise de movimento em tempo real (menos de 33ms para processar um par de imagens em um notebook AMD Turion Dual Core 2GHz). Como principal resultado, na maior parte do tempo, a f?vea m?vel permite ao rob? n?o realizar movimenta??o f?sica de seus dispositivos rob?ticos para manter uma poss?vel regi?o de interesse vis?vel nas duas imagens. Validamos o modelo proposto com resultados experimentais