[Zurück]

M. Ayromlou, M. Vincze, W. Ponweiser, M. Zillich:
"Cue Integration for Model-based Feature Tracking";
Vortrag: 27th Workshop of the Austrian Association for Pattern Recognition, Laxenburg; 05.06.2003 - 06.06.2003; in: "Vision in a Dynamic World", (2003), ISBN: 3-85403-168-8; S. 11 - 18.

Robust visual sensing is still the main bottleneck for reliable robotic systems that are capable of
acting in real world scenarios. In particular, the complexity and unpredictability of realistic and
unconstrained environments are the most limiting factors. Therefore it is the major objective of
this work to achieve robustness of vision-based tracking in natural surroundings. An approach
for model-based object tracking is introduced, which exploits cues derived from the image, the
object model and the pose of the object. Their evaluation is organized according to their
contextual relevance. Local cues (such as edge, intensity or color) are applied first to restrict
matching ambiguity at an early stage. Successively, global model cues (i.e. topological feature
relations) and pose cues (i.e. image to model feature fit) are applied to resolve ambiguities at a
higher level. Since high-level cue evaluation methods depend on the results of low-level methods,
at each level several redundant cues are evaluated to obtain more reliable performance. Results
of the successful application of cue integration are presented.

http://aleph.ub.tuwien.ac.at/F?base=tuw01&func=find-c&ccl_term=AC04404088