As mobile robotics research is gradually moving beyond problems such as 2D navigation, localization, and mapping, the ‘next big challenge’ is the understanding of 3D environments. In this context, ‘understanding’ entails the robust perception of 3D geometry, the abstraction of this geometry to a set of meaningful entities, and finally the interpretation of these entities to infer information about the state of the environment and to possibly perform manipulation or interaction.
In this line of research we focus only on data obtained from a 3D laser range finder setup. This is in parallel to the research area “Cognitive understanding of the Environment”, that focuses on the fusion of vision and range data. While visual sensors provide a very convenient way of gathering information about the environment at high rates, the information from laser range finders is more accurate and contains all necessary information for classification on geometric level.
The above figure illustrates our chain of abstraction that leads from raw point cloud data to recognized objects.
As a first step, 3D point clouds are acquired with a suitable sensing setup, in our case an actuated 2D laser range finder. In a second step, descriptive features such as normals and curvature values are computed for each of the acquired points. Using an efficient segmentation algorithm, point clouds are then abstracted into sets of segments with homogeneous curvature change. Basic instances of oversegmentation caused by partial occlusions are efficiently remedied by merging gaps that have likely been caused by occluding surfaces. Finally, segmented scans of many different objects are used to learn a vocabulary of common parts and perform Bayesian classification of new scans based on the normalized part histograms of learned object classes.
Resources:
» Selected data sets used for segmentation
Please direct questions and inquiries regarding this line of research to Klaas Klasing.
Researcher
Leading Researcher
Dirk Wollherr
Project Researcher
Klaas Klasing