Logo U. Osnabrueck Institute of Computer ScienceKnowledge-Based Systems Research Group

Martin Günther - Research


Martin Günther Research CV Publications Downloads

PhD Project

Title: Using Top-Down Information and Context in Semantic Mapping
Supervisor: Prof. Dr. Joachim Hertzberg
Abstract:

One of the classical tasks of a mobile robot is building a map of its environment. This has led to a large body of work on the problem of simultaneous localization and mapping (SLAM). Traditionally, those maps are used mainly for navigation purposes, which only requires geometric information. However, more demanding applications of mobile robots (e. g., service or rescue robots) require semantically meaningful structures in the environment to be extracted, a process known as semantic mapping. Semantic maps can be useful for a variety of tasks; we focus on planning goal-directed interaction with objects in the environment. The goal of this dissertation project is to develop a system that integrates high-level semantic knowledge into the mapping process, and uses an HTN planner to perform given tasks on the objects detected in the semantic map.

An important feature of our system is that the data flow is not solely bottom up. Our claim is that top-down semantic knowledge is essential for more basic perceptive functions. Expectations about typical size, position and spatial relations to other objects can be used to disambiguate noisy object classifications or guide an active search for undetected objects.

The proposed system consists of four layers:

  1. the object recognition layer, which uses data from a 3D laser scanner to model the environment geometry and classify objects;
  2. the local context layer, which computes probabilities of expected objects based on the context (i. e., other objects in the scene and the type of scene), using a Bayesian Network;
  3. the global conceptual layer, which represents objects, aggregates and places using a Description Logic ontology; and
  4. the planning layer, which uses an HTN planner to plan robot actions.

The Bayesian Networks in layer 2 will only be used locally and generated from the information in layer 3, possibly using the PR-OWL system. The thesis focuses on layers 24, with the perspective of integrating object recognition from another PhD project. The system will be implemented on a Kurt type robot platform.

Grants: This research is part of the PhD Programme Cognitive Science and funded by a Lichtenberg scholarship from the Ministry for Science and Culture of Lower Saxony.
 

Martin Günther Research CV Publications Downloads
Last changed: 2012-05-11