Image Understanding

The members of the research group Image Understanding work primarily in the fields of image understanding, software development for autonomous systems (robots, vehicles), computational neuroscience, and biology. An integral part of all research is the use of massively parallel programs which allow applications to be carried out as efficiently and close to real-time as possible. This aspect is particularly important in neuroscience, since neural networks are a priori parallel processing systems.

Robotics and Distributed AI

The control of autonomous systems requires considerable sensorial support. The development of these supporting components is thus the primary task of image understanding. In order to enhance the capacity and speed of most of the existing algorithms for basic image processing (filtering, stereo processing), they are re-implemented on SIMD-machines (e.g. MasPar). Symbolic pictures (e.g. semantic networks) are deployed for situation recognition (e.g. at crossroads). The networks of MIMD-machines are particularly useful.

Three mobile robot systems are available, building the basis for the investigation of cooperative performances, e.g. convoy driving or multi-agent forming in a production environment. Besides the standard equipment, one vehicle has a 6-dof manipulator, while a second vehicle has a stereo head, controllable in four axes. Another topic is the simulation of cooperations (planning and communication) of autonomous agents in order to learn about methods of self-organization in distributed computer and machine networks.

Non-linear Dynamics

In order to control the behavior of technical systems in a suitable manner, a profound knowledge about the dynamics of such systems is necessary. Therefore it is required to devise an adequate mathematical model which can be simulated and analyzed with appropriate investigation methods. On the basis of the results obtained in this way, one can derive suitable control mechanisms for the considered technical systems. The control of large scale systems, consisting of many interacting subsystems, can rely also on self-organization principles.

As a paradigm one can think about selection equations respectively coupled selection equations. These are derived from physical, chemical or biological systems which show under certain conditions the emergence of macroscopic patterns, that means temporal, spatial, spatio-temporal or even functional structures. Within the research activities of the project DZOP approaches and methods based on selection equations and hence on concepts of non-linear dynamics and self-organization are meanwhile sucessfully applied to the control of autonomous mobile robots and the solution of specific tasks in flexible transport and manufacturing systems.

In applications where such mechanisms or concepts are implemented they have to be adapted, optimized or even expanded. As a consequence, systems constructed in such a way have to be analyzed very acurate. Hence the efficient simulation and analysis of dynamical systems using adequate simulation tools is quite important. Therefore a simulation and analysis tool for dynamical systems AnT was developed by the NLD-group of the institute. This tool allows not only the simulation but also a qualitative and quantitative investigation of several classes of dynamical systems.