The appearance-based approach towards robot navigation is based on a pixel-wise comparison of images. Recent research has shown that the Euclidean distance in image space provides a robust method for robot homing, navigation along routes and topological mapping. The objective of this paper is to investigate how image data can be reduced in order to minimise the computational cost for the image distance calculation without loosing the robustness of the method. A simple 1-D scenario is used to test three different types of reduction methods: one focuses on specific and predefined image regions, one uses fractal sets, while the last is based on a stochastic process. We show that with less than 10% of the data a similar performance can be achieved with the stochastic method, which is then used on a real case study (the visual compass) to assess its performance in a real situation.
展开▼
