Robot self-localization using only view-to-view registration

Mobile robot often requires to locate itself in the work environment so as to determine how far or close it is away from the goal position. There have been a number of solutions proposed, but most of them require 3D model of the environment for referencing. However, such 3D model is often inconvenient and sometimes impossible to acquire. This work addresses how a robot could locate itself by using only one view of the environment for reference. More precisely, the robot captures from its current, unknown location an image of the surroundings, and matches this view to a reference view of the environment for determining its current location. The required view-to-view registration could be greatly aided by placing distinct, easy-to-match landmarks in various places of the work environment. However, landmarks are often required to be small to minimize their visual disturbance to the environment, and that would compromise the localization accuracy. This work focuses on a mechanism that allows correspondences be extended from some initial ones over the landmark to a bigger set which are over parts of the scene next to the landmark, thereby increasing the physical span of the correspondences and improving the localization accuracy. In a way, the mechanism allows the scene itself to be used as a huge landmark for robot localization. Synthetic data and real image experiments are shown to illustrate the performance of the mechanism