Robust long-term positioning for autonomous mobile robots is essential for many applications. In manyenvironments this task is challenging, as errors accumulate in the robot’s position estimate over time. Therobot must also build a map so that these errors can be corrected when mapped regions are re-visited; thisis known as Simultaneous Localisation and Mapping, or SLAM.Successful SLAM schemes have been demonstrated which accurately map tracks of tens of kilometres, howeverthese schemes rely on expensive sensors such as laser scanners and inertial measurement units. A moreattractive, low-cost sensor is a digital camera, which captures images that can be used to recognise wherethe robot is, and to incrementally position the robot as it moves. SLAM using a single camera is challenginghowever, and many contemporary schemes suffer complete failure in dynamic or featureless environments, orduring erratic camera motion. An additional problem, known as scale drift, is that cameras do not directlymeasure the scale of the environment, and errors in relative scale accumulate over time, introducing errorsinto the robot’s speed and position estimates.Key to a successful visual SLAM system is the ability to continue operation despite these difficulties, andto recover from positioning failure when it occurs. This thesis describes the development of such a scheme,which is known as BoWSLAM. BoWSLAM enables a robot to reliably navigate and map previously unknownenvironments, in real-time, using only a single camera.In order to position a camera in visually challenging environments, BoWSLAM combines contemporary visualSLAM techniques with four new components. Firstly, a new Bag-of-Words (BoW) scheme is developed, whichallows a robot to recognise places it has visited previously, without any prior knowledge of its environment.This BoW scheme is also used to select the best set of frames to reconstruct positions from, and to findefficient wide-baseline correspondences between many pairs of frames. Secondly, BaySAC, a new outlier-robust relative pose estimation scheme based on the popular RANSAC framework, is developed. BaySACallows the efficient computation of multiple position hypotheses for each frame. Thirdly, a graph-basedrepresentation of these position hypotheses is proposed, which enables the selection of only reliable positionestimates in the presence of gross outliers. Fourthly, as the robot explores, objects in the world are recognisedand measured. These measurements enable scale drift to be corrected. BoWSLAM is demonstrated mappinga 25 minute 2.5km trajectory through a challenging and dynamic outdoor environment in real-time, andwithout any other sensor input; considerably further than previous single camera SLAM schemes.
展开▼
