Comparing holistic and feature-based visual methods for estimating the relative pose of mobile robots

Abstract

Feature-based and holistic methods present two fundamentally different approaches to relative-pose estimation from pairs of camera images. Until now, there has been a lack of direct comparisons between these methods in the literature. This makes it difficult to evaluate their relative merits for their many applications in mobile robotics. In this work, we compare a selection of such methods in the context of an autonomous domestic cleaning robot. We find that the holistic Min-Warping method gives good and fast results. Some of the feature-based methods can provide excellent and robust results, but at much slower speeds. Other such methods also achieve high speeds, but at reduced robustness to illumination changes. We also provide novel image databases and supporting data for public use.