Learning to acquire and select useful landmarks for route following

The paper describes a prototypical system for optimal landmark acquisition and selection. Our landmark-learning approach does not require any type of environment model to be supplied to the robot in advance, and represents a step towards robots interacting with real environments. The approach is fitted and tested for the TMGA system (P. Zingaretti et al., 1998), which the authors developed for landmark tracking by adaptive, stereo template matching. Two complementary strategies, properly managed, are followed to construct a suitable subset of landmarks: the selection of the more discriminant landmarks and the selection of the landmarks that are more invariant in a neighbourhood. The robustness of the TMGA system in analysing the discriminant power of each landmark and the analysis of the disparity map and of the spatial activity maps of the stereo images are used for identifying discriminant and invariant landmarks. The experimental results show that the number of matching failures, and consequent landmark changes during the following of a route is comparable with (not much greater than) those obtained using an a-priori subset.