Localization, mapping and recognition in outdoor environments

Camera systems play a key role in mobile robotics because of their reliability in varied environments, compact size and cost effectiveness. Regardless of the assigned task, a key component of any mobile robot is the ability to localize itself in its environment. In the most general case, this means that it should be capable of mapping its environment during exploration and recognizing objects and previously visited locations in this map. Mapping and recognition together allow the robot to localize itself in its environment. This thesis contributes towards two aspects of the localization problem: loop closing and mapping. We use a monocular camera to recognize previously visited locations (loop closing) in large-scale urban environment [28]. We use GPS positioning to narrow down the search area and use a pre-trained vocabulary tree to find the best matching image in this search area. Geometric consistency is used to reject the false positive matches. We compare several vocabulary trees on outdoor urban sequences. We experiment with Hierarchical K-means [44] based trees as well as Extremely Randomized Trees [18] and compare results obtained using five different trees. Our second contribution studies a constrained version of the camera pose estimation problem, where the camera positions are known in a fixed world coordinate system, using a GPS, for example. Formally, we study the problem of estimating camera orientations from multiple views, given the positions of the viewpoints in a world coordinate system and a set of point correspondences across the views [5]. Given three or more views, the above problem has a finite number of solutions for three or more point correspondences. Given six or more views, the problem has a finite number of solutions for just two or more points. In the three-view case, we show the necessary and sufficient conditions for the three essential matrices to be consistent with a set of known baselines. We develop algorithms to impose the known baseline constraints during Structure from Motion and study their suitability under noise through experiments on synthetic data. We also conduct experiments on data from the ICCV2005 Computer Vision Contest [55] and self acquired data