Axel Steinhage

Image processing and behavior planning for intelligent vehicles

Since the potential of soft computing for driver assistance systems has been recognized, much effort has been spent in the development of appropriate techniques for robust lane detection, object classification, tracking, and representation of task relevant objects. For such systems in order to be able to perform their tasks the environment must be sensed by one or more sensors. Usually a complex processing, fusion, and interpretation of the sensor data is required and imposes a modular architecture for the overall system. In this paper, we present specific approaches considering the main components of such systems. We concentrate on image processing as the main source of relevant object information, representation and fusion of data that might arise from different sensors, and behavior planning and generation as a basis for autonomous driving. Within our system components most paradigms of soft computing are employed; in this article we focus on Kalman filtering for sensor fusion, neural field dynamics for behavior generation, and evolutionary algorithms for optimization of parts of the system.

Self-calibration based on invariant view recognition: Dynamic approach to navigation

We extend a dynamic approach of behavior generation to the representation of spatial information. Two levels of dynamics integrate dead-reckoning, dominant far from home bases, and piloting, dominant near home bases. When the view-based piloting system recognizes a home base, visual place information recalibrates the dead-reckoning system, inverting the hierarchical ordering of the two dynamic levels by time scale inversion. Reference views taken at discrete home bases are recognized invariantly under rotation of views. This process yields compass information. Continuous translational information is obtained as a neural place representation built from view correlations with a scattered set of local views. This self-calibrating cognitive map couples into a dynamics of heading direction integrating the behaviors of obstacle avoidance and target acquisition. Targets can be designated in terms of the cognitive map. We demonstrate the dynamical model in simulation.

Complex behavior by means of dynamical systems for an anthropomorphic robot

We present an architecture to generate behavior for an anthropomorphic robot. The goal is to equip the robot with the capacity to interact with a human. Motivated by the research on biological systems, our basic assumption is that the behavior to perform determines the external and internal structure of the behaving system. We describe the anthropomorphic design of our robot and present a distributed control system that generates human-like navigation and manipulation behavior. As the mathematical framework for this purpose we have developed a control system which is entirely based on dynamical systems in the form of instantiated dynamics and neural fields. We also present a dynamic scheme for the behavioral organization based on competitive dynamics.

Dynamical systems for the behavioral organization of autonomous robot navigation

We present an architecture for the behavioral organization of autonomous robots. For the example of navigation, we describe how complex behavior can be broken up into multiple elementary behaviors. The overall behavior is generated by activating and deactivating the elementary behaviors dependent on both the sensor input and the intrinsic logics of the behavioral plan needed to fulfill the task. The elementary behaviors as well as their organization into behavioral sequences are achieved by appropriately designed nonlinear dynamical systems. We show how intrinsicly discrete functionalities like counting and decision making can be realized by nonlinear dynamical systems and how these dynamics can be coupled stably and flexibly.

Nonlinear attractor dynamics: a new approach to sensor fusion

Fusing information of multiple sensor is particularly difficult if the sensor systems which provide the information have very different characteristics such as different data formats, reliabilities, signal to noise ratios, sampling rates and so on. Furthermore, the information is often provided on different levels of abstraction such as the direct sensor output in contrast to expert knowledge or a priori information. We propose a new approach to sensor fusion which accounts for these problems. The basic idea is to represent the quantity to estimate as the state variable of a nonlinear dynamical system. The sensor signals act on this dynamics by specifying attractors with limited range of influence. The dynamics relaxes into a stable state which results from the superposition of the attractors. By means of the limited attractor ranges, the dynamics automatically averages nonlinearly over corresponding sensor signal while outliers stemming form temporarily de-calibrated or erroneous sensor are discarded. Self-calibration is achieved by representing also the sensor signals as dynamical states and specifying an attractor at the position of the fused estimate. By using the unified attractor representation, abstract information can be treated in the same way as direct sensor input. Furthermore a mathematically well defined and algebraically analyzable format for dynamic sensor information on various levels of abstraction is available. We verify our concept for the example of man-machine interaction: fusing visual and odometric sensor information for the autonomous position estimation with acoustic guidance information for the target acquisition of a mobile robot.

Attractor dynamics to fuse strongly perturbed sensor data

Presents a new approach to multi-sensor fusion which is based on coupled nonlinear attractor dynamics. The state of the dynamics represents the fused estimate of a physical entity measured by multiple sensors. Each sensor-reading but also general expert knowledge about the measured system specifies a local stable fixed point (attractor) with a limited basin of attraction of the dynamics. The dynamic state variable converges to a global stable state which is the systems fused estimate. For the example of measuring the oil film thickness on seawater by means of multispectral radiometer measurements gathered during flights across a polluted area, the authors show that their approach is particularly useful for fusing multimodal strongly perturbed sensor data.

Population Coding in Cat Visual Cortex Reveals Nonlinear Interactions as Predicted by a Neural Field Model

We develop population coding ideas toward a general approach to the analysis of cortical function that operationalizes the notion of cooperativity. Neural ensemble activation distributions (population representations) are constructed over a defined stimulus parameter space, in our case the 2-dimensional retinal position of the central visual field. In contrast to classical approaches using receptive field centered stimuli the method presented here requires the stimulation of a whole cell ensemble with an identical common stimulus. The constructed activation distribution allows a quantitative investigation of activation dynamics and cooperative effects, like lateral inhibition and excitatory interaction. We simulated the data with a continuous neural network model as proposed by Wilson & Cowan [14].

A robust self-calibrating data fusion architecture

The authors present a general mathematical framework for the fusion of noisy sensor data. On the basis of the mathematical theory of dynamical systems they couple the outputs of the sensors to obtain a nonlinearly averaged overall estimate of the physical quantity to measure which automatically discards outliers from the averaging process. Drifts within the time series of single sensors can be compensated through a recalibration by the method of time scale inversion. By means of a unified way of representing information as stable states of a dynamical system it is possible to integrate different sorts of information such as expert knowledge and sensor information smoothly within the data fusion system. They verify the feasibility of their approach on the basis of simulated stochastic data sets and on the basis of data from a study in which the brightness temperature of oil films on sea water was measured. The proposed self-calibrating sensor fusion architecture extends the work they presented at IGARSS 99 in Hamburg (1999).

Multiple classifier system based on attractor dynamics

A new method is proposed for combining outputs of several classifiers. The method is based on the theory of dynamical systems. In our formulation each classifier output represents a forcelet in a phase space of the dynamical system. Depending on interactions (superposition) of forcelets the system can perform in two different regimes: non linear averaging among several classifiers outputs and selection among them. We show that the attractor dynamics method outperforms both the winner takes all algorithm and the single best classifier.

Dynamical systems for sensor fusion and classification

In this paper we show how the dynamic approach to sensor fusion, presented on IGARSS 1999 and IGARSS 2000 can be applied to the problem of classifying noisy sensor data. The idea is to use the output of the dynamic sensor fusion algorithm as input for a system of winner-takes-all dynamics in which different classes compete with each other. In this way, transitions between classes are brought about by bifurcations between stable states of a dynamical system. For the example of classifying sea ice types from SAR image data, we will show that, due to the defined time scale of these bifurcations, the dynamic approach is advantageous for classifying properties of real physical systems.