Institute for
Robotics and Process Control

Deutsch   English

Probabilistic Object Recognition

Decription of project

The recognition of a 3d scene is a very complex process due to the different analysis and interpretation algorithms for the sensor data. Furthermore, the erroneous data often cause ambiguities within the interpretation step. A robust recognition system has to integrate different information sources and extract robust features to overcome these ambiguities. Especially the recognition of freeform is challenging due to the error prone features based on differential surface properties. We solve this problem by modeling the statistical behavior of feature and object. The use of Bayes nets allows characterizing the statistical behavior and the relation between feature and object. The following Figure illustrates a Bayes net for recognition of polyhedral objects..

drseitcu_net_print_eng2.gif

The corners of a polyhedron represent the 3d features. A Bayes net models this object corners as features for the recognition process.

pd_scene.gif

pd_scene_disp.gif
Scene with polyhedral objects and results.

Fundamental for the configuration of the 3d recognition step is the automatic generation of the Bayes net that is divided into two steps. First, the Bayes net is constructed offline. Then, the observed scene is analyzed with the Bayes net and the objects within the scene are detected. The preprocessing extracts features from the CAD data in the same way the recognition do from the sensor data. The similarity of the features defines the dependencies in the Bayes net.

bkr2.gif
Feature extraction for the automatic generation of a Bayes net

Border curves are another discriminative feature. A border curve divides two surfaces and can be extracted from CAD and depth data. They are well-defined by curvature and torsion and can be decomposed into part curves. This allows the representation by a hierarchical Bayes net, where the leaves are the part curves and the root is the object. The intermediate layer represents the border curve. Coded light approaches are used to provide flexible acquisition of depth data from different views. The following applications are implemented in the workspace of a robot.

bkr3.gif
3d object recognition for polyhedral scenes


adept2.gif
3d object recognition of brake and fan belt disk


bkr5.gif
Scene with large occlusions and results.


It took 0.25s to generate this page.