Bärbel Mertsching

Data- and model-driven gaze control for an active-vision system

Models of visual attention provide a general approach to control the activities of active vision systems. We introduce a new model of attentional control that differs in important aspects from conventional ones. We divide the selection into two stages, which is more suitable for the system as well as explaining different phenomena found in natural visual attention, such as the dispute between early and late selection. The proposed model is especially designed for use in dynamic scenes. Our approach aims at modeling as much of a general active vision system as possible and designing clean interfaces for the integration of the remaining specific aspects needed in order to solve specific problems.

Combining Multiple Views and Temporal Associations for 3-D object Recognition

This article describes an architecture for the recognition of three-dimensional objects on the basis of viewer centred representations and temporal associations. Considering evidence from psychophysics, neurophysiology, as well as computer science we have decided to use a viewer centred approach for the representation of three-dimensional objects. Even though this concept quite naturally suggests utilizing the temporal order of the views for learning and recognition, this aspect is often neglected. Therefore we will pay special attention to the evaluation of the temporal information and embed it into the conceptual framework of biological findings and computational advantages. The proposed recognition system consists of four stages and includes different kinds of artificial neural networks: Preprocessing is done by a Gabor-based wavelet transform. A Dynamic Link Matching algorithm, extended by several modifications, forms the second stage. It implements recognition and learning of the view classes. The temporal order of the views is recorded by a STORE network which transforms the output for a presented sequence of views into an item- and-order coding. A subsequent Gaussian-ARTMAP architecture is used for the classification of the sequences and for their mapping onto object classes by means of supervised learning. The results achieved with this system show its capability to autonomously learn and to recognize considerably similar objects. Furthermore the given examples illustrate the benefits for object recognition stemming from the utilization of the temporal context. Ambiguous views become manageable and a higher degree of robustness against misclassifications can be accomplished.

Integration of Static and Dynamic Scene Features Guiding Visual Attention

This paper presents a visual attention module driven by static and dynamic scene features controlling the gaze shifts of an active vision system. A preattentive processing unit computes several static features, like orientation and color, and a dynamic feature, motion. We distinguish two further processing modes of our active vision system: the hypothesis validation mode and the tracking mode. In the hypothesis validation mode the bottom-up static features and top-down information of the presence of an object are combined to guide the recognition process. The preattentive dynamic feature analysis represents an alert system. By the presence of motion it interrupts the hypothesis validation mode and triggers the tracking mode. Several experimental results are presented.

Playing Domino: A Case Study for an Active Vision System

We introduce a mobile robot playing at dominoes. The robot is equipped with a pan-tilt unit and a CCD camera and is solely guided by the vision system. Besides the mobile robot the overall system is distributed on a general-purpose workstation. The communication is realized by three radio links transmitting the video signal and two serial data streams for the pan-tilt and the robot controller respectively. Two client/server connections enable a bidirectional data exchange between the image processing software and the robot control and navigation software. The protocol includes requests for certain actions to be performed by the opposite module and the transmission of domino coordinates. The robot identifies the dominoes, collects them, and deposits them in form of a straight chain where two adjacent domino halves possess the same digit. Several iterations of a perception action cycle have to be executed to fulfill this task: The first cycle starts with a visual exploration of the robots environment and a subsequent explicit foveation of detected domino point clusters. After a coordinate transformation the dominoes are identified by template matching followed by a comparison with a knowledge base to eliminate ambiguities. The positions of the dominoes are registered into the global map space of the robot which position is continuously updated while it is moving. When the robot has reached a so-called virtual point in front of a specified domino the robot requests a control view towards this domino to verify its position. After feedback from the vision system the robot grips the target, moves to the appropriate position in front of the already formed chain, again verifies its position, and deposits the domino. Afterwards the robot returns to its initial position where the cycle starts again with the selection of the next target domino.

Perceptual grouping in grey-level images by combination of Gabor filtering and tensor voting

We present the extension of a perceptual grouping method known as tensor voting to the application on grey-level images. In addition to formerly used inputs consisting of binary images or edgel maps, we introduce the use of local orientation tensors which are computed from a set of Gabor filters applied to the input image. This approach not only yields oriented input tokens but also the locations of junctions as input to the perceptual grouping. We show how this extension can smoothly be embedded into the tensor voting framework and demonstrate the method on example images.

Application of the Tensor Voting Technique for Perceptual Grouping to Grey-Level Images

We show how the perceptual grouping method known as tensor voting can be applied to grey-level images by introducing the use of local orientation tensors computed from a set of Gabor filters. While inputs formerly consisted of binary images or sparse edgel maps, our extension yields oriented input tokens and the locations of junctions as input to the perceptual grouping. In order to handle dense input maps, the tensor voting framework is extended by the introduction of grouping fields with inhibitory regions. Results of the method are demonstrated on example images.

Design and Implementation of an Accelerated Gabor Filter Bank Using Parallel Hardware

In computer vision, images are often preprocessed by the so-called Gabor transform. Using a Gabor filter bank, an image can be decomposed into orientational components lying in a specified frequency range. This biologically motivated decomposition simplifies higher level image processing like extraction of contours or pattern recognition. However, the IEEE floating-point implementation of this filter is too slow for real-time image-processing, especially if mobile applications with limited resources are targeted. This paper describes how this can be overcome by a hardware-implementation of the filter algorithm. The actual implementation is preceded by an analysis of the algorithm analyzing the effects of reduced-accuracy calculus and the possibility of parallelizing the process. The target device is a Xilinx Virtex FPGA which resides on a PCI rapid-prototyping board.

Opponent Color Processing Based on Neural Models

In this paper we present a new opponent color system which imitates some of the known color processing neural cells established by electrophysiological recordings. We describe the benefits of this system to image processing tasks. The opponent color model is embedded in an active vision system to improve the systems fixation and recognition capabilities. This is done by removing illumination effects to some degree and by evaluating the resulting color differences. Experimental results are presented.

Architecture of a 3D-simulation environment for active vision systems and mobile robots

Working with active vision-systems, users often have to simulate particular environments in which they can undertake experiments with virtual sensors. In such artificial environments all parameter settings have to be fully controllable. Thus a simulation framework based on a component-oriented concept will be proposed in order to facilitate the integration of arbitrary simulation modules, e.g. particular sensors, environmental artifacts and even additional platforms for robots. The introduced system aims at enabling visual simulation state-monitoring. It also provides interfaces which are identical to those of hardware platforms already in use. This ensures that existing active vision or robotic applications can use the simulator in the same way. Although it is not intended to replace the real system completely, this application does represent an important addition to the system for both research and teaching purposes.

Application of the tensor voting technique for perceptual grouping to grey-level images: quantitative evaluation

This paper presents a quantitative evaluation of the application of the perceptual grouping method known as tensor voting to grey-level images. For that purpose, we have introduced the use of local orientation tensors computed from a set of Gabor filters. While inputs formerly consisted of binary images or sparse edgel maps, we use oriented input tokens and the locations of junctions from images as input to the perceptual grouping. Here, we introduce a benchmark test to estimate the precision of our method with regards to angular and positional error. Results on these test images show that the computation of the tensorial input tokens is highly precise and robust against noise. Both aspects arc further improved by the subsequent grouping process.