One key application in mobile robotics is the creation of geometric maps using data gathered with range sensors in indoor environments. These maps are usually used for navigation and represent free and occupied spaces. However, whenever the robots are designed to interact with humans, it seems necessary to extend these representations of the environment to improve the human-robot interaction. In this work, we present an approach to extend indoor laser-based maps with semantic terms like "corridor", "room", "chair", "table", etc, used to annotated different places and objects in 2D or 3D maps. The approach uses an Associative Markov Network (AMN) which follows the concept of collective classification in the sense that the labeling of a data point in the space is partly influenced by the labeling of its neighboring points. In addition, the approach is relational as it learns dependencies between the particular object classes. Thus, the relations between the object classes can be modeled from the training data. AMNs classify the points in a map using a set of features representing these points. In this work, we show how to choose these features in the different cases of 2D and 3D laser scans. Experimental results obtained in simulation and with real robots demonstrate the effectiveness of our approach in various indoor environments.
Bibtex
@InProceedings{triebel2007gfkl,
title = {Relational Learning in Mobile Robotics: An Application to Semantic Labeling of Objects in {2D} and {3D} Environment Maps},
author = {Rudolph Triebel and Oscar Martinez Mozos and Wolfram Burgard},
booktitle = {Annual Conference of the German Classification Society on Data Analysis, Machine Learning, and Applications},
address = {Freiburg, Germany},
year = {2007},
url = {http://www.informatik.uni-freiburg.de/~omartine/publications/triebel2007gfki.html},
}