András Barta

Basic course in control theory

Course material of basic control theory has been overviewed and updated recently at the Faculty of Electrical Engineering and Informatics, BME. The paper describes shortly the contents and the teaching methods of the course. In the theoretical material the concept of the Youla parametrization has been introduced which gives a new insight into controller design. MATLAB exercises support understanding and interactively applying the theoretical knowledge. New lecture notes were written both for the theoretical material and for the MATLAB laboratory exercises. The course web-site provides the actual information and gives also additional learning materials.

Controlling the Visual Attention with Layered Dual Pyramid Architecture

Visual attention is a fast developing area of the pattern recognition field. The area has been drawn the attention of the research community, because the theory has solid biophysical support and it is capable of bringing together the various concepts of pattern recognition. The paper presents a network architecture that is capable of providing bottom up, top-down and lateral information flow. The framework to accomplish the three-directional information flow is the Layered Dual Pyramid (LDP) architecture.

Adaptive Multi-scale Vision System for Robotic Applications

The paper proposes a system for adaptive multi-scale processing of visual information. The system consists of several components: a hierarchical dynamic graph structure, a network inference algorithm and for low level processing a wavelet based disk transform. The presented system implements a two level data structure. For the high level processing a dynamic tree structure and a Bayesian inference algorithm is used. Every node contains not only position but also scale and rotation information. At the low level it uses hierarchical structure of quadratic spline wavelet image bases

STRUCTURAL OBJECT RECOGNITION BY PROBABILISTIC FEEDBACK

This paper investigates, how Bayes networks can be applied to hierarchical object representation. It provides a hierarchical graph definition an a recursive algorithm to convert the objects to graph representations and also to reconstruct the image of the objects from its graph. The structural complexity of the objects is calculated to get compact representations. The application of graphs for object recognition is an old idea, but lately it came into the limelight again. Graph structures with the help of probability models turned out to be powerfool tools. They are called graphical modells. A graphical model can capture and store the model description and also provide a computational background. Bayesian or bilief networks encodes probabilistic relationship among variables or objects. A probability is assigned to every node of the graph and the edges provide information about their dependencies. This probability represents the subjective belief or knowlage about the node. This paper investigates, how Bayes networks can be applied to hierarchical object representation. It provides a hierarchical graph definition an a recursive algorithm to convert the objects to graph representations and also to reconstruct the image of the objects from its graph. The structural complexity of the objects is calculated to get compact representations. In this paper we present only the main ideas behind our algorithm, the exact theoretical description is not given. The definition of the graph structure and

Controller Design by Direct Numerical Optimization

Abstract The available computer power recently increased so fast that problems that previously can not be handled are now solvable by straight numerical methods. The research attempts to solve control design problems without the added design expertise. The paper overviews some of the optimization algorithms that can be used in practical applications. First it shows how to solve a problem with the application of nonlinear optimization algorithms, then a simulation is presented to demonstrate how the different design objectives effects the systems behavior. The effect of parameter change is also examined.

Model Creation for Mobile Robots Using 2-d Data and Featuresx

Abstract The most critical part of the programming of an autonomous mobile robot is its model creation. In a practical application it is not sufficient to design only the robot’s path, the surrounding objects have to be manipulated also, which requires an internal model of the surrounding objects. This paper presents a method for model creation. It is based on the combination of a global optimization algorithm and feature detection. It examines how different type of data can be build into the optimization process. The paper evaluates the algorithm in a two-dimensional simulated environment. It shows how to identify predefined objects using a single grayscale image.

Model Creation for Factory Floor Robot

Abstract Recent developments in mobile robot designs made it possible to apply them in a real world situation. In a practical application it is not sufficient to design only the robots path, the surrounding objects have to be manipulated also. The calculation of the movement of the robot requires an internal model of the surrounding objects. This paper presents a method to build such a model. The algorithm works with the two-dimensional projections of the objects. The proposed method builds the workspace from predefined objects. To do this it uses a global optimization algorithm.

Identifying Objects for Factory Transportation Robot

Abstract One fast growing area of autonomous mobile robots is factory floor transportation robots. They perform simple object manipulation or transportation functions. The calculation of the movement of the robot requires an internal model of the surrounding objects. This paper investigates how a 2-D model can be created using a single gray-scale picture. The first part of the paper explores the possibilities of converting the picture to a form that can be used for model identification. The second part examines how this preprocessed data can be used to identify predefined objects. A simulated factory floor is used to evaluate the algorithm.