We demonstrate the use of our active object recognition algorithms in a mechanical assembly task. The algorithms are used to classify and estimate the pose of parts of the assembly in different stable rest positions and automatically re-position the camera if the class or pose of an object is ambiguous in a given image. Multiple object views are used in determining both the final object class and pose estimate. The FSTs are analyzed off-line to determine the camera positions that best resolve ambiguities. We also describe methods for rejecting untrained objects and adding new parts to an existing set of FSTs using a new feature update method.

An intelligent robot is a remarkably useful combination of a manipulator, sensors and controls. The current use of these machines in outer space, medicine, hazardous materials, defense applications and industry is being pursued with vigor but little funding. In factory automation such robotics machines can improve productivity, increase product quality and improve competitiveness. The computer and the robot have both been developed during recent times. The intelligent robot combines both technologies and requires a thorough understanding and knowledge of mechatronics. In honor of the new millennium, this paper will present a discussion of futuristic trends and predictions. However, in keeping with technical tradition, a new technique for 'Follow the Leader' will also be presented in the hope of it becoming a new, useful and non-obvious technique.

In the future, interaction between humans and personal roots will become increasingly important as robots will, more and more, operate as assistants in our everyday life. Because of this, there is a need for a convenient, flexible, and general-purpose technique that we can use to interact with robots. Moreover, the same technique should also be usable when we interact with embedded systems in smart environments. In this paper, we will describe a technique that allows us to use a single simple handheld control device to interact not only with personal robots, but also with ubiquitous embedded systems. When a new system, whether a mobile robot or a VCR, is encountered, the control device downloads a mobile code from the system and executes it. The mobile code then sues the services provided by the control device to create a virtual user interface that we can use to interact with the particular syste. Our technique is flexible, simple, adaptive, and open. The technique draws much of its flexibility and simplicity from the mobile code. Adaptivity comes from the fact that the control device only needs minimal knowledge about each particular system. In addition, the technique does not place any restrictions on the type of mobile code that can be used. We will describe the architecture of the CUES system that utilizes our technique. We will also describe the architecture of the SMAF system, our test bed mobile code execution environment used in CUES. In addition, we will present a virtual user interface for a mobile robot that we can use to control the robot and to monitor its status information. The interface also operates as a terminal that we can use to access remote information in the Internet.

Innovative technology excels by realizing extraordinary solutions for well-known problems. The grating-optical correlation measurement technology promotes Opto-Electronics into Opto², and Opto³, and Optoⁿ-Electronics enabling precise non-contact, non-slip acquisition of data from length and speed information, to information data about directions of motion, distances, and multiple distances in 1, 2 and 3 directions all the way up to image preprocessing on vehicles, robots, production machines, and aides for blind individuals. This paper describes the technological steps of these developments.

In this paper, we describe a visual surveillance system for evaluating the audience's reaction in meeting sessions. The system, which works in real-time, can recognize and evaluate the reaction of the audience. It is mainly composed of three subsystems. The first subsystem is a face detection and head motion segmentation system which is used to detect the face from complex background and segment the head motion into different units, with each unit including different information about the audience's reaction. The second subsystem is a gesture and pose recognition syste which can recognize the gestures and the poses of human head. The third subsystem is an evaluation system which is used to evaluate the reaction of the audience by using the recognition result in the second system. Our system has been tested for determining the concentration of the audience in meeting sessions and the experimental results are good.

This paper presents an efficient method to search a target- object image appearing on an unknown scene image. The target-object is affine-transformed from the original model image, that is, translated, rotated and scaled. The Gabor transform is used to obtain the spectrum information from both the model and the scene images. The spectrum information of the model has the characteristic that the spectrum plane response regularly corresponding to the rotation and scale-change in real space. Using the characteristic, target-object can be correctly detected and the pose is also calculated by spectrum matching. Also the few Gabor functions can cover the whole frequency spectrum of the original image with less interference, very few numbers of features, that is, the Gabor-expansion coefficients, can represent the target-object. It result in the highly efficient calculation.

This paper present a method to reduce the calculation costs of object pose measurements. They are estimated by matching L-shaped line-segments in 3D object models with those in 2D object images using a P3P solution. The L segments have the minimum pieces of information for the estimation, and, therefore, it produces a large amount of L segments both in the 3D object models and in the 2D object images, resulting in enormous correspondences, and P3P calculations. To solve the problem, we propose a strategy enabling to select fewer L segments: as an evaluation function, we utilizes an average of the estimation errors when observing the L segments from representative points on a geodesic dome. Furthermore, we deduce a useful approximate formula for L segments having various shapes.

Pose characterization refers to the unique representation of an object's pose through a set of image features. In this regard, one of the prerequisite for using visual images in representing an object's pose is a set of view-dependent image features that is unique to a single position and orientation of the object relative to the camera. Here, we propose an enhancement to the global feature extraction method. It utilizes an active illumination that projects a set of orthogonal grid patterns on the object. The deformations of the grid pattern incident on the complex target object carried important information about the surface structure of the target. This grid-encoded image can be characterized using global descriptors. Results indicate that this illumination increases the sensitivity of the global features to changes in object pose.

To control the position and movement of an end-effector on the tip of a multi-joint robot arm is known to include a kind of redundant problem. Although the end-effector is set its position by each angle of the joints, the angle of each joint cannot be uniquely determined by the position of the end-effector. Each of infinite number of different sets of joint angles usually represent the same position of the end- effector. This paper describes how to control the angle of each joint to move its end-effector from a starting point to an ending point on an X-Y plane preferably. We first separate standpoints into two to define the preferable movement; 1) the standpoint of the end-effector, and 2) the standpoint of the joints. Then, we define multiple objective functions from each standpoint. Finally, we formulate the problem into a multi-purpose programming problem. We apply a genetic algorithm to solve this problem and obtain satisfied solutions, which have a smooth movement of the end-effector and less rotation of the joints. This paper is suggestive that the approach described here can easily be extended to a problem with a multi-joint robot arm in a 3D space, and also to a problem with obstacles between starting and ending points.

A new concept for visually guided manipulator control is introduced. It eliminates the need for a calibration of the manipulator as well as of the vision system and comprises an automatic adaptation to changing parameters. A key point of the concept is the achievement of a complex and elaborate desired goal by activating an appropriate sequence of rather simple elementary behaviors. Contrary to conventual stereo vision methods it uses a calibration-free camera system and allows a direct transition from image coordinates to motion control commands of a robot. By this approach, the abstract coordinate transformations have been avoided, instead, image data are used directly to control the behavior of the root, or the interactions of the robot with physical objects. Thus, it makes knowledge of many hard-to-measure optical and mechanical system unnecessary; moreover, it lends itself to the realization of learning and adaptive robots. The concept has been successfully realized and tested in real-world experiments with a visually guided calibration-free 5 degree of freedom manipulator invovling the grasping of various objects with nearly any shape in arbitrary position in the robot 3D work space.

This work demonstrates a vision-based control technique that does not require robot or vision system calibration. There are two distinct advantages: first, the approach is generic and can be applied to a variety of systems; second, calibration is unnecessary after a reconfiguration or disturbance to the robotic workcell. It has the potential to provide a low-cost, low-maintenance automation solution for unstructured industries and environments. The robot end- effector tracks a moving target using a novel dynamic quasi- Newton control was formulated in the image plane and on-line Jacobian estimation using either a dynamic Broyden's method or a dynamic recursive least squares algorithm. Experimental results demonstrate convergent and stable control of an uncalibrated manipulator tracking a moving target. The method is shown to be robust to system reconfiguration such as modifications to the position and orientation of the camera.

A Neuro-fuzzy control method for navigation of an Autonomous Guided Vehicle robot is described. Robot navigation is defined as the guiding of a mobile robot to a desired destination or along a desired path in an environment characterized by as terrain and a set of distinct objects, such as obstacles and landmarks. The autonomous navigate ability and road following precision are mainly influenced by its control strategy and real-time control performance. Neural network and fuzzy logic control techniques can improve real-time control performance for mobile robot due to its high robustness and error-tolerance ability. For a mobile robot to navigate automatically and rapidly, an important factor is to identify and classify mobile robots' currently perceptual environment. In this paper, a new approach of the current perceptual environment feature identification and classification, which are based on the analysis of the classifying neural network and the Neuro- fuzzy algorithm, is presented. The significance of this work lies in the development of a new method for mobile robot navigation.

Motion control is one of the most critical factors in the design of a robot. The purpose of this paper is to describe the research for applying motion control principles for a mobile robot systems design, which is on going at the University of Cincinnati Robotics Center. The mobile robot was constructed during the 1998-1999 academic year, and called BEARCAT II. Its design has inherited many features of its predecessor, BEARCAT II, such as vision guidance, sonar detection and digital control. In addition, BEARCAT II achieved many innovative motion control features as rotating sonar, zero turning radius, current control loop, and multi- level controller. This paper will focus on the motion control design, development and programming for the vehicle steering control and rotating sonar systems. The systems have been constructed and tested at the 1999 International Ground Robotics Competition with the BEARCAT II running an obstacle course for 153.5 feet and finishing fourth in the competition. The significance of this work is in the increased understanding of robot control and the potential application of autonomous guided vehicle technology for industry, defense and medicine.

An autonomous guided vehicle is a multi-sensor mobile robot. The sensors of a multi-sensor robot system are characteristically complex and diverse. They supply observations, which are often difficult to compare or aggregate directly. To make efficient use of the sensor information, the capabilities of each sensor must be modeled to extract information form the environment. For this goal, a probability model of ultrasonic sensor (PMUS) is presented in this paper. The model provides a means of distributing decision making and integrating diverse opinions. Also, the paper illustrates that a series of performance factors affect the probability model as parameters. PMUS could be extended to other sensor as members of the multi-sensor team. Moreover, the sensor probability model explored is suitable for all multi-sensor mobile robots. It should provide a quantitative ability for analysis of sensor performance, and allow the development of robust decision procedures for integrating sensor information. The theoretical sensor model presented is a first step in understanding and expanding the performance of ultrasound systems. The significance of this paper lies in the theoretical integration of sensory information from the probabilistic point of view.

The purpose of this paper is to describe a high-level path planning logic, which processes the data from a vision system and an ultrasonic obstacle avoidance system and steers an autonomous mobile robot between obstacles. The test bed was an autonomous root built at University of Cincinnati, and this logic was tested and debugged on this machine. Attempts have already been made to incorporate fuzzy system on a similar robot, and this paper extends them to take advantage of the robot's ZTR capability. Using the integrated vision syste, the vehicle senses its location and orientation. A rotating ultrasonic sensor is used to map the location and size of possible obstacles. With these inputs the fuzzy logic controls the speed and the steering decisions of the robot. With the incorporation of this logic, it has been observed that Bearcat II has been very successful in avoiding obstacles very well. This was achieved in the Ground Robotics Competition conducted by the AUVS in June 1999, where it travelled a distance of 154 feet in a 10ft. wide path ridden with obstacles. This logic proved to be a significant contributing factor in this feat of Bearcat II.

In this paper we describe how a robot detects its state with respect to a goal using only visual information and consequently how it learns to reach that goal navigating in a free environment. The state detection is carried out using a feed-forward neural network, with several multiple input and output units, trained using the quickprop method that is an optimized variant of the back-propagation algorithm. The color images captured by the on-board camera of the robot, are a color coded and pre-processed to construct a robust set of inputs to the net, taking account of the trade-off between the dimension of the input set and the loss of information in the image. The simple goal-reaching behavior, finally, is learned using a reinforcement learning algorithm with which the robot associates a proper action to each detected state. To speed up this learning phase an initial state-action mapping is learned in simulation. Starting from this basic knowledge, the real robot will continue to learn the optimal actions for reaching the goal since new unexplored situation can occur in the real environment. The results obtained experimenting this approach on the real robot Nomad200 are described in the paper.

Diffractive 3D phase gratings of spherical scatterers dense in hexagonal packing geometry represent adaptively tunable 4D-spatiotemporal filters with trichromatic resonance in visible spectrum. They are described in the (lambda) - chromatic and the reciprocal (nu) -aspects by reciprocal geometric translations of the lightlike Pythagoras theorem, and by the direction cosine for double cones. The most elementary resonance condition in the lightlike Pythagoras theorem is given by the transformation of the grating constants g_x, g_y, g_z of the hexagonal 3D grating to (lambda) _h1h2h3 equals (lambda) ₁₁₁ with cos (alpha) equals 0.5. Through normalization of the chromaticity in the von Laue-interferences to (lambda) ₁₁₁, the (nu) (lambda) equals (lambda) _h1h2h3/(lambda) ₁₁₁-factor of phase velocity becomes the crucial resonance factor, the 'regulating device' of the spatiotemporal interaction between 3D grating and light, space and time. In the reciprocal space equal/unequal weights and times in spectral metrics result at positions of interference maxima defined by hyperbolas and circles. A database becomes built up by optical interference for trichromatic image preprocessing, motion detection in vector space, multiple range data analysis, patchwide multiple correlations in the spatial frequency spectrum, etc.

In this paper we prose the application of the codebook computed by the Self Organizing Map as a smoothing filter, the QV Bayesian Filter, for the preprocessing of the image sequences. The optical flow is then robustly and efficiently computed over the filtered imags applying a correlation approach at the pixel level.

In 1990, Mount Holyoke College began an AI project centered around a mobile robot the students have named 'Susan B.' One of the research threads in this project is what has been called the 'Where am I.' problem, namely, the study of how to determine the robot's location using Computer Vision. This paper describes our current approach to this problem.

A method for locating the doors in an image is presented. The method integrates an edge map with color information known about the doors to provide better results than methods based purely on color or edges. The Hough transform is used to find lines and then a set of heuristics is used to find possible door regions. The door regions are tested by determining whether the region is the correct color. The combination of color and edge data allowed the system to successfully identify doors within a complex environment at different scales, orientations, lighting conditions and using different cameras.

3D color histograms are introduced as an effective means of object recognition. No globally optimal set of color histogram parameters is known, and the choice of data-set specific parameters is far from obvious due to the size of the search space involved. Evolution Strategies (ES), a form of Evolutionary Computation, are introduced as a method of optimizing histogram parameters specific to a known data set. An ES is implemented on a 22-object, 110 image database, and a 93 percent recognition rate achieved, a significant improvement over the 86 percent recognition rate of standard histogram axes. The results demonstrate the efficacy of ES and underscore the importance of the assumptions that histogram-based recognition methods are built upon.

Surface color measurement is of importance in a very wide range of industrial applications including paint, paper, printing, photography, textiles, plastics and so on. For a demanding color measurements spectral approach is often needed. One can measure a color spectrum with a spectrophotometer using calibrated standard samples as a reference. Because it is impossible to define absolute color values of a sample, we always work with approximations. The human eye can perceive color difference as small as 0.5 CIELAB units and thus distinguish millions of colors. This 0.5 unit difference should be a goal for the precise color measurements. This limit is not a problem if we only want to measure the color difference of two samples, but if we want to know in a same time exact color coordinate values accuracy problems arise. The values of two instruments can be astonishingly different. The accuracy of the instrument used in color measurement may depend on various errors such as photometric non-linearity, wavelength error, integrating sphere dark level error, integrating sphere error in both specular included and specular excluded modes. Thus the correction formulas should be used to get more accurate results. Another question is how many channels i.e. wavelengths we are using to measure a spectrum. It is obvious that the sampling interval should be short to get more precise results. Furthermore, the result we get is always compromise of measuring time, conditions and cost. Sometimes we have to use portable syste or the shape and the size of samples makes it impossible to use sensitive equipment. In this study a small set of calibrated color tiles measured with the Perkin Elmer Lamda 18 and the Minolta CM-2002 spectrophotometers are compared. In the paper we explain the typical error sources of spectral color measurements, and show which are the accuracy demands a good colorimeter should have.

This work present one method aimed to individualization and recognition of vial signs in route and city. It is based fundamentally on the identification by means of color and form of the vial sing, located in the border of the route or street in city, and then recognition. To do so the obtained RGB image is processed, carrying out diverse filtrates in the sequence of input image, or intensifying the colors of the same ones otherwise, recognizing their silhouette and then segmenting the sign and comparing the symbology of them with the previously stored and classified database.

We apply a model of texture segmentation using multiple spatially and spectrally localized filters, known as Gabor filters, to the analysis of texture and effect regions found on wooden boards. Specifically we present a method to find an optimal set of parameters for a given 2D object detection method. The method uses banks of Gabor filters to limit the rang of spatial frequencies, where mutually distinct textures differ significantly in their dominant characterizing frequencies. By encoding images into multiple narrow spatial frequency and orientation channels a local classification of texture regions can be achieved. Unlike other methods applying Gabor filters, we do not use a full Gabor transform, but use feature selection techniques to maximize discrimination. The selection method uses a genetic algorithm to optimize various parameters of the system including Gabor weights, and the parameters of morphological pre-processing. We demonstrate the applicability of the method to the task of classifying wooden textures, and report experimental results using the proposed method.

Classification of real-time x-ray images of randomly oriented touching pistachio nuts is discussed. The ultimate objective is the development of a subsystem for automated non-invasive detection of defective product items on a conveyor belt. We discuss the use of clustering and how it is vital to achieve useful classification. New clustering methods using class identify and new cluster classes are advanced and shown to be of use for this application. Radial basis function neural net classifiers are emphasized. We expect our results to be of use for other classifiers and applications.

This paper presents a robot based surface inspection application for measuring mould surface in foundries. The surface measuring its performed using a robot with six degrees of freedom equipped with a laser-triangulation-based distance sensor. The measuring process is divided into four phases: sensor calibration, calibration of the mould location, surface inspection measurements and measurement analysis. In the sensor calibration phase, tool correction is calculated by using algorithms based on least square estimation with Newton iteration. Equations derived for calibration uncertainty estimation are verified using Monte Carlo simulations. The real covariance of sensor tool correction has been obtained with test measurements. Calibration of the mould location is based on the same estimation principle as was used in sensor calibration. Total uncertainty of the measuring system is obtained by transforming all separate uncertainties into one total uncertainty covariance. Spatial uncertainties are expressed and manipulated in the form of covariance matrices. The volume of the uncertainty ellipsoid in 3D space is calculated for each sensor calibration and mould location calibration. For comparison, the goodness of the measurement model is evaluated also by the condition number of the Jacobian matrix.

This work is a part of ongoing research in the area of automatic visual inspection systems for real-time detection of fabric defects. The study aims to extend and evaluate the application of the joint space/spatial-frequency approach represented by the use of Gabor elementary functions for inspecting intricate jacquard patterns. It assesses the utility of multiresolution properties of Gabor-filters and the need for adaptive selection and integration of appropriate resolution levels of the image pyramid. The choice of the appropriate levels takes into consideration characteristics of the potential defects.

This paper deals with an important task within forensic science - the automatic comparison of bullets for the purpose of firearm identification. Bullets bear groove- shaped marks that can be though of as a kind of 'fingerprint' of the firearm on their circumferential surface. To accomplish the comparison task, mainly the fine grooves on the bullet surface are of interests.

A reconstructed scene in virtual reality typically consists of millions of triangles.Data are heterogeneous and consist not only of geometric coordinates but also of multi-modal data. The latter requires more complex calculations and very high-speed graphics. Due to the large amount of data, displaying and analyzing these 3D models require new methods. This paper present an innovative method to analyze multi-model models using a 2D-quincunx wavelet analysis. The algorithm is composed of three processes. First, a set of range images is captured from various viewpoints surrounding the object of interest. In addition, a set of multi-modal images is acquired. Then, a multi-resolution analysis based on the quincunx wavelet transform is performed. The multi- resolution analysis allows extraction of multi-resolution detail areas. These areas of details are projected back onto the surface of the initial model. Detail areas are marked onto the model and constitute another modality. Finally, a mesh simplification is performed to reduce data that are not marked as detail. This approach can be applied to any 3D models containing multi-modal information in order to allow fast rendering and manipulation. This method also allows 3D data de-noising.

The determination of point correspondences between range images is used in computer vision for range image registration and object recognition. The use of a spin image as a feature for matching has had considerable success in object recognition. However, in registration, refinement by iterative methods has been required. This paper present a method of determining the surface geometry in a local region surrounding the point. The technique is developed for range images which have little movement between viewpoints, and which consists of only several profiles each. The method involves fitting surface patches to the surfaces of the two successive views, creating spin-image features at a few points of each patch in one view, and determining the best match of features on the previous reference view using a localized interpolating search. The sets of corresponding points of the two successive range views are then used directly to compute the registration transformation between views. This computation effectively refines the corresponding by minimizing the residual errors. The technique is demonstrated using a pair of synthetic range views, derived from a range image of an object with a free- form surface.

In this paper a sequence of algorithms for image feature point detection and tracking as well as Euclidian reconstruction of rigid 3D objects from point correspondence is presented. For fully automatic point feature detection, gray value images are processed. High-curvature points on contours, i.e. contour elements with locally maximal curvature, are tracked using a normalized correlation algorithm. High-curvature points that could be tracked in a sequence of more than three images are used as point feature that are eligible for reconstruction. Since the general way to obtain the epipolar, projective, and Euclidian geometries from point feature correspondence is already solved, here the emphasis is on the performance of the algorithms in the presence of noise. Kanatani's epipolar geometry estimation method is improved and this is experimentally validated. Regarding Bougnoux's Euclidian geometry estimation method, the initial linear solution is now obtained with less uncertainty and the non-linear minimization does no longer converge to a hidden solution. Experimental results are given to assess the system performance.

This paper present a new concept for 3D free-form surface registration and object recognition using a new surface representation scheme. This representation scheme captures the 3D curvature information of any free-form surface and encodes it into a 2D image corresponding to a certain point on the surface. This image is unique for this point and is independent from the object translation or orientation in space. For this reason we called this image 'Surface Point Signature'. This scheme can be used as s global representation of the surface as well as a local one and also in a scale independent surface matching. It performs faster registration than existing registration approaches.

This paper introduces a fast and efficient indexing approach for both 2D and 3D model-based object recognition in the presence of rotation, translation, and scale variations of objects. The indexing entries are computed after preprocessing the data by Haar wavelet decomposition. The scheme is based on a unified image feature detection approach based on Zernike moments. A set of low level features, e.g. high precision edges, gray level corners, are estimated by a set of orthogonal Zernike moments, calculated locally around every image point. A high dimensional, highly descriptive indexing entries are then calculated based on the correlation of these local features and employed for fast access to the model database to generate hypotheses. A list of the most candidate models is then presented by evaluating the hypotheses. Experimental results are included to demonstrate the effectiveness of the proposed indexing approach.

This paper describes an algorithm for the automatic segmentation and representation of surface structures and non-uniformities in an industrial setting. The automatic image processing and analysis algorithm is developed as part of a complete on-line web characterization system of a paper making process at the wet end. The goal is to: (1) link certain types of structures on the surface of the web to known machine parameter values, and (2) find the connection between detected structures at the beginning of the line and defects seen on the final product. Images of the pulp mixture, carried by a fast moving table, are obtained using a stroboscopic light and a CCD camera. This characterization algorithm succeeded where conventional contrast and edge detection techniques failed due to a poorly controlled environment. The images obtained have poor contrast and contain noise caused by a variety of sources.

Indexing is an important aspect of video database management. Video indexing involves the analysis of video sequences, which is a computationally intensive process. However, effective management of digital video requires robust indexing techniques. The main purpose of our proposed video segmentation is twofold. Firstly, we develop an algorithm that identifies camera shot boundary. The approach is based on the use of combination of color histograms and block-based technique. Next, each temporal segment is represented by a color reference frame which specifies the shot similarities and which is used in the constitution of scenes. Experimental results using a variety of videos selected in the corpus of the French Audiovisual National Institute are presented to demonstrate the effectiveness of performing shot detection, the content characterization of shots and the scene constitution.

We are interested in the development of an efficient computational visual system to study dynamic fluids and in this paper we present a methodology to detect patterns in flow-like images. These patterns are called in the Dynamical Systems theory critical points or single points. This method, like other previous methodologies presented in related literature, involves four steps: estimation of flow directions; detection of critical point candidates; matching of critical points; and critical points description according to the Dynamical Systems theory. To achieve the critical point class, node, saddle or center point, we use templates to match the patterns. The last step of the method in necessary only if we wish a numerical representation of the patterns. We ran the method over several real images of some fluid flow classes and present the results.

Heated debates were taking place a few decades ago between the proponents of digital and analog methods in information. Technology have resulted in unequivocal triumph of the former. However, some serious technological problems confronting the world civilization on the threshold of the new millennium, such as Y2K and computer network vulnerability, probably spring from this one-sided approach. Dire consequences of problems of this kind can be alleviated through learning from the nature.

This paper addresses the problem of architecture optimization, when implementing an image matching primitive in reconfigurable circuits. Circuit spatial organization is optimized in terms of processing time, and circuit volume, in order to suit well for real time on board applications. This optimized adaptive spatial and scalable organization of the (mu) PD circuit dedicated to image matching reduces by one order the spatial and temporal performance, without altering the quality of matching. The (mu) PD circuit has been validated with the minimal 22 elementary cells architecture with Xilinx 4010 XL circuit working at 12 MHz and occupying 92 percent of the circuit CLB. It performs the pyramidal 256 X 256 image matching in less than 1 s.

Human visual system is properly suited for reliable and adequate volumetric perception of natural environment. Volumetric data flows coming from the outer physical space are easily acquired, transferred and processed by eye-brain system in real time. This relates also to the animals which use different complicate mechanisms of optical volumetric data acquisition and can navigate safely at high speeds. On the contrary machine vision systems utilizing currently the stereoscopic effect in attempt to achieve volumetric data presentation are very slow, bulky and in a way inelegantly devised. The stereoscopy itself seems can hardly organize the adequate, real time volumetric robot vision.

Problem of pattern recognition can be interpreted as a problem of acceptance of optimal decision under conditions of uncertainty, caused by absence of the complete and authentic information about a recognized object and its features. The unique adequate method of solving of pattern recognition problem in the conditions of uncertainty is the decisions making by the whole set of available heterogeneous information, taking into account a significance and reliability of each of considered feature and their interrelation. Usually the solution of pattern recognition problem is reduced to the task of minimization of distance from an image of the object up to the standard image of the class of objects. In this paper we offer and review the possible approach to generalization of the Mahalnobis metrics, based on properties of fuzzy number in L-R form. The results of the experimental comparison of the effectiveness of pattern recognition using the considered set of fuzzy features and criteria are discussed.

We present here a review of scientific theoretical researches performed during last five years at our Institute. Specific modern radiative transfer approaches and techniques. This is just that is meant as 'nontraditional' features. All these features formed a basis for evaluating somewhat surprising, but easily physically treated properties of images provided by active vision systems operating through a turbid medium. Specific topics are: (i) effects of fine backscattering pattern of coarse aerosols, such as fog or cloud droplets, on optical interference and the computational consequences of these effects; (ii) image contrast and ultimate visibility range of targets with different reflective properties and; in particular, possible improving of the visibility range of targets with different reflective properties and; in particular, possible improving of the visibility of a target as it is sinking into a turbid medium; (iii) imaging of non-Lambertian objects and peculiarities in their images; (iv) applications to assessing visibility quality of a car driver under poor weather conditions and some ways to optimize the visibility and to enhance the ultimate visibility range. We succeed in treating all these topics by rather simple analytical expressions requiring no sophisticated software to be dealt with.

This paper shows how graph and diagrammatic structures could evolve into the more abstract ones that carry knowledge about the original in the form of relations and hierarchies. They can play a role of context, or 'measurement device', giving the ability to analyze. Such derivative structures can drive processes in distributed networks, like firing a spatio-temporal pattern, creating another structure, etc., therefore, performing top-bottom algorithms. In the mid-80 PDP group came closely to the problem of knowledge representation by distributed graphs called 'schemata'. Their models were based on the neural networks. I argue that the actual level of problems that neural network can solve is lower than required for knowledge representation. The 'hardware unit' of intelligence could rather be a 'neural assembly' that combines both discrete and continuous features, and is able to perform both diagrammatic and graphs operations, being the basis of intelligence. The representation of such 'neural assembly' as a fuzzy logic unit with active kernel and less active fuzzy boundaries is proposed. Such units can make logical operations spatially and temporarily, acting on a diagrammatic manner. Connections between the set of such activated units make spatio-temporal pattern really looking like a set of nodes connected via links on more abstract level. The article shows that image understanding is the are where derivative structures play important role, making images and scene self-describing without any work description, and truly invariant to any transformations. That opens the way to the new technologies in computer vision and image databases.