Vision-based place recognition for mobile robots

In this thesis, methods for representation, recognition, and prediction of environmental features such as landmarks and places using monocular image data for mobile robot navigation are developed. Landmarks are often used as a basis for mobile robot navigation. We consider the problem of automatically selecting from a set of 3D features the subset which is most likely to be recognized from noisy monocular image data and is least likely to be confused with any of the other groups of features. Assuming perspective projection, real valued recognition functions are constructed for a set of features. The value returned from such functions are invariant to changes of viewpoint and can be evaluated directly from image measurements without prior knowledge of the position and orientation of the camera. With image noise, the recognition function no longer evaluates to a constant value. Because of the possibility of false matches, a Bayes detector is used to determine the optimal range of values of the recognition function that will be accepted. The collection of features with the lowest Bayes cost is selected as the most distinguishable landmark. We show implementation results for real 3D objects.; Central to many mobile robot navigation systems is a graph-based representation or map of space whose nodes are places and whose arcs are actions or behaviors that can be executed to move a robot between places. To automatically construct such a representation as a robot explores, it is necessary to recognize when the robot has revisited a place. Here the notion of place is extended beyond a neighborhood of a point in space to a large region (e.g. a room), and a method for place recognition from a single image is presented. Places are represented using a set of images of the place, taken from different viewpoints. An algorithm to automatically choose the most salient groups of features to model a place from image data is described. Places are recognized by establishing correspondences of point and line segment features through constrained search and geometric invariants. The constrained motion of a mobile robot reduces the combinatorics of the matching process, leading to fast, effective place recognition.; While following a path using vision-based robot navigation, visual landmarks can be tracked and their image locations can be used to control robot motion. In general, tracking algorithms take advantage of the fact that features only move a short distance between images, and so only a small area of the image needs to be searched. During navigation, landmarks must be located and acquired as they come into view. The ability of the robot to predict the locations of the landmark features from image data will reduce the search area and simplify the tracking process. We discuss methods for direct image-based prediction of line segment and straight line features for a mobile system operating on a planar surface. Preliminary experimental results suggest that image-based prediction can by performed efficiently and with sufficient accuracy to ensure robust acquisition of navigational landmarks. Some issues, improvements and extensions to the method are discussed for all the algorithms developed.