Jan C. Vorbrüggen

Distortion invariant object recognition in the dynamic link architecture

An object recognition system based on the dynamic link architecture, an extension to classical artificial neural networks (ANNs), is presented. The dynamic link architecture exploits correlations in the fine-scale temporal structure of cellular signals to group neurons dynamically into higher-order entities. These entities represent a rich structure and can code for high-level objects. To demonstrate the capabilities of the dynamic link architecture, a program was implemented that can recognize human faces and other objects from video images. Memorized objects are represented by sparse graphs, whose vertices are labeled by a multiresolution description in terms of a local power spectrum, and whose edges are labeled by geometrical distance vectors. Object recognition can be formulated as elastic graph matching, which is performed here by stochastic optimization of a matching cost function. The implementation on a transputer network achieved recognition of human faces and office objects from gray-level camera images. The performance of the program is evaluated by a statistical analysis of recognition results from a portrait gallery comprising images of 87 persons. >

GripSee: A Gesture-Controlled Robot for Object Perception and Manipulation

We have designed a research platform for a perceptually guided robot, which also serves as a demonstrator for a coming generation of service robots. In order to operate semi-autonomously, these require a capacity for learning about their environment and tasks, and will have to interact directly with their human operators. Thus, they must be supplied with skills in the fields of human-computer interaction, vision, and manipulation. GripSee is able to autonomously grasp and manipulate objects on a table in front of it. The choice of object, the grip to be used, and the desired final position are indicated by an operator using hand gestures. Grasping is performed similar to human behavior: the object is first fixated, then its form, size, orientation, and position are determined, a grip is planned, and finally the object is grasped, moved to a new position, and released. As a final example for useful autonomous behavior we show how the calibration of the robots image-to-world coordinate transform can be learned from experience, thus making detailed and unstable calibration of this important subsystem superfluous. The integration concepts developed at our institute have led to a flexible library of robot skills that can be easily recombined for a variety of useful behaviors.

GripSee: A Robot for Visually-Guided Grasping

We have designed an anthropomorphic robot system at our institute as a research platform and demonstrator for the next generation of service robots. GripSee is a visually guided robot which is endowed with a number of skills based on different neural network architectures. The skills include the interpretation of human gestures, localization and recognition of objects, planning and generation of grasping movements, and automatic calibration of the eye-hand coordination. This paper gives an overview of the system and reports on our experiences in applying diverse neural network architectures in a real robot.

Applying Dynamic Link Matching to Object Recognition in Real World Images

We apply the dynamic link matching algorithm to object recognition in gray level images. The algorithm is able to map from one view of an object to different-e. g., translated, rotated, or mirror-reflected—views, being at the same time tolerant of small distortions. A sparse representation (10%) of the image data is used as a boundary condition for a self-organizing mechanism which performs the object match within a modest number of iterations (~102). The mechanism can be derived from local neural dynamics [1].

Parallelverarbeitung in Hardware: Optimierung numerischer Routinen auf dem T800

Die Mikroprozessor-Familie von INMOS, Transputer genannt, vereinigt einen RISC-Prozessor mit zusatzlicher Hardware auf dem Chip sowie direkter Unterstutzung im Instruktionssatz fur ein Parallelverarbeitungssystem nach dem CSP-Modell. Der T800 fugt dem noch eine Flieskommaeinheit hinzu, die ebenfalls auf dem Prozessorchip integriert ist. Dadurch konnte der Flaschenhals in der Kommunikation zwischen dem Prozessor und der Flieskommaeinheit beseitigt werden, der konventionelle numerische Koprozessoren in ihrer Leistung beschrankt [1].

Bilderkennung mit dynamischen Neuronennetzen

Es wird ein Objekterkennungssystem beschrieben, um damit die Fahigkeiten der Dynamic Link Architecture zu illustrieren. Bei dieser handelt es sich um ein neues System fur neuronales Rechnen. Objekte werden durch dunne Graphen reprasentiert. Die Knoten der Graphen werden mit lokalen Leistungsspektren (Morlet Jets) etikettiert, die Kanten mit Abstandsvektoren. Objekte werden durch Graphenvergleich erkannt. Die Ahnlichkeit zwischen Objektgraphen und Bildgraphen wird durch einen Diffusionsprozes in der Bildebene optimiert, wobei gleichzeitig die einzelnen Jet-Ahnlichkeiten maximiert und die metrische Graphenverzerrung minimiert werden. Wir haben das System auf einem Transputernetzwerk implementiert, um seine Parallelisierbarkeit zu demonstrieren. Es ist sehr erfolgreich im Erkennen von menschlichen Gesichtern anhand von frei aufgenommenen Kamerabildern.