In order for a mobile robot to perform its assigned tasks, it often requires a representation of its environment, a knowledge of how to navigate in its environment, and a method for determining its position in the environment. These problems have been characterised by the three fundamental questions of mobile robotics, that is ``Where am I?'', ``Where am I going?'' and ``How can I get there?''. This thesis is concerned principally with a method for answering the first question, that of position estimation, which is commonly referred to as localisation.
A naive approach to robot localisation is to use odometers or accelerometers to measure the displacements of the robot. This approach, known as dead reckoning, is subject to errors due to external factors beyond the robot's control, such as wheel slippage, or collisions. More importantly, dead reckoning errors increase without bound unless the robot employs sensor feedback in order to recalibrate its position estimate.
A key issue in developing a solution to the localisation problem is that of domain dependence. The majority of localisation methods are constructed based on explicit and/or implicit assumptions about the environment . In the context of machine vision, for example, many techniques extract domain-dependent features from the image, such as straight lines or corners - features which may not be present or stable outside of structured office or industrial environments. Hence, a goal of the method presented here is to achieve domain independence. We achieve this goal by learning the features or landmarks which are useful for the particular domain under consideration. While moving to a new domain will always require retraining (that is, exploration), the methods presented here are general in the sense that the same off-line training and on-line estimation methods can be employed in a new domain with little or no alteration.
In addition to domain dependence, solutions to the problem of robot localisation are necessarily sensor-dependent. For example, a wide range of solutions have been proposed which use sonar sensor data. Given the sparse two-dimensional information that a sonar sensor provides, these particular solutions are unlikely to be easily applicable to the dense three dimensional image data obtained from a laser range-finder or stereo cameras. Furthermore, the geometric interpretation of two- or three- dimensional range data obtained from sonar, stereo cameras, or laser range-finders is subject to instabilities and issues of non-invertibility. In addition, each sensor has its own set of strengths and limitations. Sonar sensors are inherently noisy, whereas laser range-finders and stereo cameras can be prohibitively expensive or difficult to calibrate. For all of these reasons, we have chosen to use a single digital camera for localisation. The camera is a sensor which is at once inexpensive, and provides large quantities of data at low computational cost, possessing relatively a low signal-to-noise ratio.