Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Fast 3D cube vision for real-world systems

: Smieja, F.

urn:nbn:de:0011-b-733257 (1.3 MByte PDF)
MD5 Fingerprint: 2e63b64a2dc7207388fdecfe52fe87db
Created on: 07.08.2002

Sankt Augustin: GMD Forschungszentrum Informationstechnik, 1998, 39 pp.
GMD Report, 5
Study, Electronic Publication
Fraunhofer AIS ( IAIS) ()
edge detection; reflection; team; 3D-synthesis

The step from robot simulation to a real-world machine is a big one, and arguably the most daunting aspect for anyone performing this step is the sudden need to introduce real perception into their system. The most typical and enticing way to enrich sensory information is to open the robot's eyes to the visual spectrum. The visual world is however one of arbitrary complexity and to hope for a general method for performing any particular aspect of visual recognition would be unrealistic. Therefore one must introduce constraints in the form of assumptions about the scene interesting to the robot and for the task to be performed. This requirement becomes most clear when speed is of the essence, as it typically is for robotic applications. This paper describes a method that makes a number of assumptions about the scene in order to generate a fast cuboid model (around five cubes per second) of cuboids interesting for our robot, JANUS. The 3D information is obtained through the use of two cameras mounted on a common movable head. This paper concerns itself also with the embedding and extension of such an algorithm in a reflective team architecture.

Contents S.v-vi
1 Introduction S.1
2 Algorithm overview S.1-2
- 2.1 Assumptions S.1-2
- 2.2 Advantages gained from the assumptions S.2
- 2.3 The detection process S.2
3 Blob extraction S.3
4 Edge detection S.4-15
- 4.1 Basic edge detection S.4-10
- 4.2 Line formation S.10-12
- 4.3 Finding vertices S.12-14
- 4.4 Guided edge detection S.14-15
5 3D-synthesis S.16-19
- 5.1 Calibrated cameras S.16-17
- 5.2 Generating a 3D point S.17
- 5.3 Corresponding points S.17
- 5.4 Generating a 3D cube model S.18-19
6 Re nements to the algorithm S.20-21
- 6.1 Incomplete 2D projections S.20
- 6.2 Errors and estimates S.20-21
7 Performance tests S.22-27
- 7.1 Measurement of correctness S.22
- 7.2 Standard cube results S.22
- 7.3 Dependence on orientation S.22-24
- 7.4 Dependence on lighting S.24
- 7.5 Dependence on blob size S.24
- 7.6 Dependence on cube color S.24-27
- 7.7 Dependence on shadows S.28
8 Parameter estimation S.28
- 8.1 General parameters S.28
- 8.2 Speci c parameters S.29
9 Algorithm limitations and the concept of re ective teams S.29-33
- 9.1 Re ective team architecture S.30-32
- 9.2 3D cube algorithm as a non-re ective team S.32
- 9.3 Introducing re ection S.32-33
- 9.4 Self-assessment S.34
10 Conclusions S.34
References S.35