Glass confidence maps building based on neural networks using laser range-finders for mobile robots

Abstract

In this paper, we propose a method to classify glass and non-glass objects and build glass confidence maps for indoor mobile robots using laser range-finders (LRFs). The glass confidence map is aimed to improve robot localization systems' robustness and accuracy in glass environments. For most LRF-based localization systems, objects are assumed to be detectable from all incident angles, which is true for non-reflective and non-transparent objects, like walls. However, glass can only be detected by LRFs in certain incident angles. This glass detection failure decreases robots' localization accuracy. Exhibiting glass' position in the map and taking its detection failure into consideration can increase the localization accuracy. We propose the usage of a neural network to classify glass and non-glass objects, with LRF's measured intensity, distance and incident angles as inputs. We verified our method experimentally, and experimental results show that our method can successfully distinguish glass from non-glass objects and accurately construct a glass confidence map with high confidence.