New large class filters are used for the automatic recognition of characters. A special set of multi-level multi-filters (referred to as "iconic filters") is described for use in the super-class recognition problem. Dif-ferent types of iconic filters are considered and the results of such tests are reported. System performance under simulated nonideal conditions is detailed. The observed behavior of iconic filters is quantified, and an initial explanation for the effect of a large number of training set images is provided. New solutions are advanced, and the the utility of iconic filters for large multi-class problems is demonstrated. This is the first data to be reported on the use of multi-level multi-filters for a large class pattern recognition problem.

The visual features most important for object recognition are those having to do with the shape of an object. In one approach to recognition The image is decomposed an image into a set of relatively simple shapes (parts) and predicates that describe these parts and the relationships between them. Recognition is a matter of matching this parts-relations description to some parts-relations model in a large database of models. Matching is computationally intensive; unless the parts-relations representation is organized for efficient indexing, recognition times get intractably long. We review general requirements for such a representation as proposed by Marr and Nishihara[1]. Their prescription leaves open general questions that must be answered in any particular implementation. We offer some solutions for a simple domain of 2-D sticklike objects and point out some aspects of the implementa-tion that might be useful for shape indexing in other domains.

An approach to 3-D surface estimation is introduced where surfaces are modelled as patches of primitive parameterized surfaces, and these parameters are estimated from images taken by cameras in two or more positions. Appropriate parameterized patches for modelling manufactured objects include planar, cylindrical, spherical and unrestricted quadric patches. Any parameterized function can be used. A complex 3-D object can be described as a collection of such patches. For more irregular surfaces such as outdoor terrain, we use a collection of primitive patches linked together as a stochastic process. More specifically, in this paper we use planar primitive patches, and describe their dependence by Markov Random Fields. We then treat 3-D terrain height estimation as stochastic pro-cess estimation given two or more image data sets. To the extent that the models used are appropriate, this should result in the most accurate surface height estimation, object recognition for objects such as boulders, crevices, mountains, etc, occulded surface height estimation, and other information extraction. A simple geometric explanation is given for the estimation algorithm. This paper is one aspect of a new Bayesian approach to stereo vision as presented in [4,5,6].

The potential use of the Hyerplane Projection (HP) method of solving a system of linear equations in the determination of synthetic discriminant functions is investigated. A systematic way of solving a system of linear equations based on its consistency/inconsistency rather than the rank of the coefficient matrix is offered. Some examples showing the promises and limitations of this approach are presented. Such an approach can also be used to find the rank of any real matrix, and can easily be extended to deal with complex quantities in general.

This paper is concerned with the problem of clustering a set of independently drawn unlabbled samples into homogeneous clusters. Each sample is a data vector drawn from a class with a probability distribution with known structure, and parametrized by a parameter vector a. No priori knowledge is assumed of the parameter set, or the number of available classes. For large data records the MLE is approximately Gaussian. We exploit this asymptotic property, and regard the the MLE set as the unlabelled samples. The clustering problem becomes then that of the identification of a mixture of Gaussian clusteries. We use the generalized mixture likelihood as the clustering metric. This metric was found to perform very close to the Bayes classifier while avoiding the computational burdens associated with the mixture likelihood maximization. The paper also addresses the problem of sparse data and deviation from normality assumption. The method is illustrated for the textured image segmentation.

A position, rotation, and scale invariant pattern recognition algorithm which is robust in non-random cluttered scenes is described. The algorithm's ability to address background noise problems found in target recognition work is examined. Simulation results using realistic aircraft scenes are also presented.

The purpose of this paper is two-fold. First, it is intended to provide some preliminary results of a character recognition scheme which has foundations in on-going neural network architecture modeling, and secondly, to apply some of the neural network results in a real application area where thirty years of effort has had little effect on providing the machine an ability to recognize distorted objects within the same object class. It is the author's belief that the time is ripe to start applying in ernest the results of over twenty years of effort in neural modeling to some of the more difficult problems which seem so hard to solve by conventional means. The character recognition scheme proposed utilizes a preprocessing stage which performs a 2-dimensional Walsh transform of an input cartesian image field, then sequency filters this spectrum into three feature bands. Various features are then extracted and organized into three sets of feature vectors. These vector patterns that are stored and recalled associatively. Two possible associative neural memory models are proposed for further investigation. The first being an outer-product linear matrix associative memory with a threshold function controlling the strength of the output pattern (similar to Kohonen's crosscorrelation approach [1]). The second approach is based upon a modified version of Grossberg's neural architecture [2] which provides better self-organizing properties due to its adaptive nature. Preliminary results of the sequency filtering and feature extraction preprocessing stage and discussion about the use of the proposed neural architectures is included.

This article describes the method which emploies the membership's principle of fuzzy mathematics in automatic recognition of the type of automobiles. The article derives the membership's function, introduces the recognition system and the recognition classifier using the micro computer Z-80. Enumerates the classification of automobile's type. Draws up recognition program. The results of recognition may display instantly on the cathode ray tube or printout.

The invariant recognition of forms is important for many tasks. The purpose of this paper is to consider algebraic and moment invariants for perspective transformations. These are important because every lens system induces a perspective transformation. The approach consists of considering the non-linear perspective transformation in a higher dimensional, homogenous space. In homogeneous space the perspective transformation is linear and algebraic invariant theory may be used to determine absolute algebraic and moment invariants. The cross ratio is a well known perspective invariant. New moment invariants corresponding to the perspective transformation are derived. Examples are presented to demonstrate the theoretical approach. The significance of this work lies in the importance of invariant recognition for both human and machine vision.

One of the main problems in image segmentation is the proper selection of thresholds. In this paper a procedure for optimum threshold selection is described. Threshold levels are chosen in such a way that the expected value of the overall segmentation error is minimized. Segmentation error is defined as some function of the error between the input image and the segmentation output. A set of equations is derived for the parameters of the segmentation and a solution for them is indicated. Experimental results are presented. Results of this technique have potential significance in many areas of computer vision and robotics.

An important problem in computer vision and image understanding is how to establish a correspondence between points in multiple images of an object. In this paper we illustrate how shape can be used in establishing point correspondences for objects bounded by smooth surfaces. This approach uses local geometric surface features in 3-dimensional world space as opposed to features in the intensity space of an image, as is done elsewhere in the image understanding literature.

Moment invariants have proven to be a useful tool for pattern recognition due to their invariance under translation, scaling and rotation. However the calculation of the moments requires a substantial amount of computer time and this has limited their usefulness in the past. The calculation of all moments up to two dimensions has been implemented solely in hardware for a high resolution image. This reduces the calculation time significantly and makes them suitable for real time applications.

The rapid and accurate inspection of microelectronic structures is necessary to increase the yield in semiconductor manufacturing. Machine inspection using black and white vision is increasingly being used to automate this function. Recently, color vision has become available as a tool for machine vision. In this paper we present the use of color vision for the measurement of oxide thickness on a silicon wafer. We have achieved a resolution of less than 30 Angstroms in thickness discrimination. The system can also be used to monitor uniformity of oxide thickness across the surface of a wafer. The sensitivity and stability of the method, along with the factors that affect the two, are examined in order to ensure long term stability and repeatability.

In this study, the correspondence problem is analyzed for two different camera-arrangement models based on their geometrical and topological properties. Emphasis is given to the model where the two images (image planes), acquired by moving a camera along its optical axis, are parallel to one another. The advantage of this arrangement over the one where the two images are coplanar is demonstrated, and a correspondence process regarding the motion-in-depth model is proposed.

Edge detection in machine vision usually consists of filtering the image with a set of circularly symmetric and/or even and odd symmetric oriented filters covering a range of spatial scales. The filters' responses at each point in the image are then thresholded either before or after being combined in some manner. Selecting functions to combine responses of filters with differing spatial scales, orientations, and symmetries is a major problem with this type of approach, as is choosing appropriate thresholds. Additionally, the computational burden has rendered the approach unfit for most practical image processing systems at this time. A new "constrained matched filter" algorithm is presented which addresses these problems. At each pixel, the algorithm computes a consistency measure and forms a template based on simple measurements of changes in intensity gradient in a small neighbourhood. Consistency is a measure of the localization of gradient changes within the neighbourhood. The location of a possible edge pixel, which need not coincide with the template center, is determined. The template is cross-correlated with the image, and the result is accumulated in an output image at the edge-pixel location previously found. The result image may be thresholded to generate a "line drawing" showing the locations of lines, step edges and roof edges.

To optimize edge detection with the familiar Laplacian-of-Gaussian (v2G) operator, it has become common to implement this operator with a large (30-by-30 or so) digital convolution mask followed by some interpolation of the processed data to determine the zero crossings that locate edges. It is generally recognized that this large mask causes substantial blur-ring of fine detail. We show that the spatial detail can be improved by a factor of about four with either the Wiener-v2G filter or an image-plane processor. The Wiener-v2G filter minimizes the image-gathering degradations if the scene statistics are (at least approxi-mately) known and also serves as an interpolator to determine the desired zero crossings di-rectly. The image-plane processor forms the v2G response by properly combining the opti-cal design of the image-gathering system with a minimal 3-by-3 lateral-inhibitory processing mask. This approach, which is suggested by (Marr's model of) early processing in human vision, also reduces data processing by about two orders of magnitude and data transmission by up to an order of magnitude.

The understanding of line drawings of scenes of Legoland (a blocks world with object boundaries belonging to only three directions) can be obtained by exploiting properties of planar surfaces and of perspective projection. The normal versors of pannels are first recovered, using the rule of maximal visibility and properties of vanishing points and of horizon lines. The usual labelling of edges can be derived from the knowledge of normal versors.

This paper uses a special approach to recognized the block pictures of Chinese characters by comparing their stochastic sectionalgrams which are obtained from original samples. In order to calculate the risk the absolute value of the difference between the image-occurr-ence probabilities of corresponding quanta in two sectionalgrams is summed. One of these two sectionalgrams is derived from the input pattern and the other from the prototype pattern. The input pattern recognition rate is inverse proportional to the value of the risk. The Markovian dynamic programming is used in this paper to check the risk. Further more, the circular layer code approach is used in the recognition system such that the recognition of input object is independent of the object's input direction. By following the different type of quanta expression, there are two Markovian dynamic programming algorithm presented in this paper to recognize the circular layer code pattern.

Recent research in motion analysis is concerned with finding a stable algorithm to recover 3-D motion and structure for practical use. We claim that any robust algorithm must improve the 3-D solution adaptively over time. As an initial step toward this goal, we investigate the behavior of the optical flow field over a short time period. This approach yields a new method for recovering 3-D motion and structure. The surface of the object is assumed to be locally planar. The 3-D velocity vectors are assumed to be constant or varying slowly. We prove the deformation parameter of the first kind, or equivalently the first order flow approximation (in space and time), is sufficient to recover 3-D rigid body motion and the surface normal. We also demonstrate, through sensitivity analysis and experiments, that 3-D inferences can be made reliably.

Stereo imaging, a branch of artificial vision, is used to determine objects' locations in three dimensions. The primary computational problem in stereo imaging is the identification of corresponding locations in two images, i.e. the matching problem. Traditionally, matching methods have been of three types: brute force (determining intensity or edge correlation between "sliding windows"), shape based (connectivity analysis yielding geometric features), or region segmentation based on classifiers (texture). We present initial results from the development of two alternate methods of image matching: color and varying camera baseline separation. While monochrome stereopsis must depend primarily on determination and matching of geometric features, color provides an added dimension that can either support geometric feature matching or alleviate stringent geometric matching criteria. Color features, extracted by use of varying color filters and signal separators, are used in the determination of corresponding objects between two images. In addition, in conventional stereo vision systems, the baseline separation of the two cameras is fixed, yielding a compromise between the resolution achievable and the ease with which features can be matched. A varying baseline strategy preserves the correspondence of paired features in each image, while allowing an increase in resolution.

Over the last two or three years the adequacy of the classical 2nd difference zero-crossing approach to edge detection in computer vision has been questioned. It has been shown that zero-crossing analysis can lose data or introduce artifacts in some situations. Although the 2nd difference approach is largely driven by the fact that human vision appears to process in 2nd difference space the present authors have concluded that perceptual behaviour is best explained as being controlled by 1st differences of luminance. Starting with a basic set of building blocks from our interpretation of human early visual function a computer simulation has been assembled which derives very simply and directly both 1st and 2nd difference maps of the input scene, both of which exhibit super-resolution capabilities. This paper describes the simple processing sequence which permits such maps to be simply derived. It is shown that the 1st difference map is free from ambiguities and contains both profile and shading data, whilst the 2nd difference map, when cued from the 1st difference map, can yield unambiguous and highly accurate motion and stereo information by simple analysis of pairs of frames.

Some properties of the motion of rigid objects are presented. The motion of a rigid object can be recovered if three vanishing points and two additional points are put in correspondance in two successive frames. Moreover it is possible to obtain relevant features of the motion of wiewed objects by the analysis of singular points of the 2-D velocity field. The main result is that pure rotation gives always nodes and as a consequence singular points associated to focuses imply rotations.

Six different algorithms (Single Differencing, Double Differencing, Linear Interpolated Differencing, Parabolic Interpolated Differencing, Spatial Differencing and Spatial Filtering) are investigated to judge their ability to track subpixel targets in moving background and additive noise. This investigation used a set of computer generated imagery for the targets, background and the additive noise. Based on this extensive simulation study, guidelines are provided about the selection of the algorithms.

This paper presents a first step in the definition and construction of a vision system to recognize three-dimensional objects whose surface are composed of planar patches and patches of quadrics of revolution. A large portion of man-made objects can be either modeled exactly, or at least well approximated by a number of patches of such surfaces. The problems addressed in this paper are three-dimensional surface recognition and parameter extraction from noisy depth maps viewing one surface. The basis for the surface recognition and parameter estimation is the curvature information that can be extracted from polynomial approximations to the range data. The curvature properties are derived directly from the Weingarten map, a well-known concept from classical differential geometry. The actual process of recognition and parameter extraction is framed as a set of stacked parameter space transforms that compute measures of confidence in an iterative refinement, or multi-level relaxation, scheme. This scheme resulted from experiments with connectionist visual recognition systems. The measure of confidence associates a measure with a hypothesis that a parameterization (corresponding to a surface) actually exists in the scene.

Whether in an office, a warehouse or a home, the mobile robot must often work in a cluttered environment; although the basic layout of the environment may be known in advance, the nature and placement of objects within the environment will generally be unknown. Thus the intelligent mobile robot must be able to sense its environment with a vision system and it must be able to analyse multiple views to construct 3-d models of the objects it encounters. Since this analysis results in a heavy computational load, it is important to minimize the number of views and to use a planner to dynamically select a minimal set of vantage viewpoints. This paper discusses an approach to this general problem and describes a prototype system for a mobile intelligent robot which can construct 3-d models from planned sequential views. The principal components of this system are: (1) decomposition of a framed view into its components and the construction of partial 3-d descriptions of the view, (2) matching of the known environment to the partial 3-d descriptions of the view, (3) matching of partial descriptions of bodies derived from the current view with partial models constructed from previous views, (4) identification of new information in the current view and use of the information to update the models, (5) identification of unknown parts of partially constructed body models so that further viewpoints can be planned, (6) construction of a partial map of the scene and updating with each successive view, (7) selection of new viewpoints to maximize the information returned by a planner, (8) use of an expert system to convert the original boundary representations of the bodies to a new Constructive Solid Geometry-Extended Enhanced Spherical Image (CSG-EESI) representation to facilitate the recovery of structural information. Although the complete prototype system has not been implemented, its key components have been implemented and tested.

Autonomous mobile robots must be capable of real-time recognition of moving objects in a time-varying scene. The motion of rigid objects can be analyzed in terms of edge features undergo-ing translation and rotation in the two-dimensional image. These features could be manipulated to match a fixed 3-D model. However, in a goal-driven system, where a priori knowledge of the 3-D model exists, it is more natural to manipulate the model before matching it to the edge features. We could speed up the 3-D to 2-D matching process by precomputing a range of possible orientations for the 3-D model. The difficulty is that one cannot predict all of the possible global changes of an object as it is moving in the scene. Our solution is to constrain the problem by looking at local changes that result from moving edge-features. The concept is borrowed from natural vision, where the visual field is divided into small regions and each region is analyzed by a column of orientation sensitive line segment operators. Each orientation operator has a preset maximal sensitivity to lines of a specific angle. Initially we assume that the 3-D model of a moving object has already been given. As an object rotates in the scene, the result is a spatial sequence of edges that activates adjacent columns. Heuristics to speed-up the matching of a 3-D model to a rotating or translating object can be derived from the local communication between neighboring columns of the moving edge features. The distance between successively activated columns can be used to measure how smoothly the edge features are undergoing motion. A slow moving object will produce a response in only a few successively adjacent columns and be interpreted as stepping slowly across the scene. A rapidly translating or rotating object will activate widely separated columns and be processed as undergoing discontinuous jumps. Similarly, in humans, when the velocity of key features in the visual field are either too low or too high, objects are perceived to be jumping instead of undergoing smooth motion [1].

A pulse-driven learning network can be applied to any problem where adaptive behavior (i.e., the ability to adjust behavior to situations where a priori solutions are not known) is important. The pulse-driven learning network approach is different from other connectionist techniques in the way communication occurs between nodes. Since other connectionist techniques allow communication to occur in a continuum fashion, solutions at each compute cycle exist only when the system is in an equilibrium state. Not only is this a very computationally intensive process, but false solutions are also possible. The learning network does not have either of these problems because communication between nodes is in the form of a pulse and the correction solution is extracted from the network in as few as ten pulses from the input nodes. The results presented herein demonstrate the ability of a pulse-driven learning network to exhibit learning from association, learning from reward/punishment for simple problems and the existence of a stable solution for solving a complex problem.

The basic configuration space method has been successfully implemented in a real time collision avoidance algorithm. The program allows a user to determine a collision free path through obstacles in the path. The program is integrated with a offline programming utility, and permits the safe utilization of a robot work cell. This research is a step towards understanding the safe operation of industrial robots.

Operating principles of three-dimensional (3D) imaging sensors are discussed and various sensor implementation techniques are reviewed. A synopsis of the relative merits of 3D imaging sensors with respect to that of other imaging sensor types is presented. An implementation technique which uses a modulated CW laser source and which has resolution versatility is described and used to illustrate applications diversity. Resulting imagery from laboratory and field measurements provides x,y,z resolution varying from 0.3 m to 0.05 mm. Illustrations presented include: (1) a system for robotic guidance for parts acquisition and inspection; (2) imagery from the 3D vision sensor utilized on the DARPA Autonomous Land Vehicle (ALV); and (3) the 3D vision sensor used on the DARPA Adaptive Suspension Vehicle (ASV). In each application, the 3D features utilized and the resulting data processing are discussed.

Establishing a correspondence between a stored model and an object in the world is a valuable process. This relationship can be performed in a "global" manner -- employing correlation techniques -- or in a "piecewise" manner -- where component matching is used. In this paper we describe the merits of establishing a piecewise correspondence between a stored model, which is a complete, viewpoint independent description, and the viewpoint dependent description that is obtained from sensing an object in the world. Piecewise correspondence is the matching of object components to model components in a unique manner. In general there will exist object components for which there are no stored model components, and model components for which there are no visible object components. This matching problem can be expressed as a Double subgraph isomorphism, where we search for the maximal subset that is common to both the model graph and the object graph and thus determine a partial match. The determination of a piecewise correspondence makes certain Image understanding tasks possible since it is known which object components are missing. Their location and type may then be hypothesized by the image understanding system. For features missing due to image processing limitations, the location of the feature in the image is determined, and the image is reexamined in an expectation-driven manner with more sensitive thresholds. If the feature is missing due to occlusion, then the sensor is repositioned so that the hypothesis can be tested from a new viewpoint. Sensor repositioning is computed first in the model-centered framework and then transformed via the established correspondence to the object-centered framework. Sensor repositioning may also serve to disambiguate object matching when a poor initial view is given. The ability to advise feature extraction routines to reprocess a section of an image, or to redirect the sensor view for hypothesis verification and disambiguation are all steps toward Image understanding. These capabilities are driven by the establishment of a piecewise correspondence between a viewed scene and a stored model. Presented in this paper is a computer vision system which performs these tasks and which will soon be incorporated into a full robotic system.

We propose a modeling system for generic objects in order to recognize different objects from the same category with only one generic model. The representation consists of a prototype, represented by parts and their configuration. Parts are modeled by superquadric volumetric primitives which are combined via Boolean operations to form objects. Variations between objects within a category are described by allowable changes in structure and shape deformations of prototypical parts. Each prototypical part and relation has a set of associated features that can be recognized in the images. These features are used for selecting models from the model data base. The selected hypothetical models are then verified on the geometric level by deforming the prototype in allowable ways to match the data. We base our design of the modeling system upon the current psychological theories of categorization and of human visual perception.

An algorithm is presented for merging three-dimensional symbolic descriptions of static polyhedral scenes obtained from multiple viewpoints. The 3-D descriptions are assumed to be partial because of occlusions. The merging algorithm treats the topology of the descriptions (i.e., connections between vertices, edges, and faces) separately from the geometry (i.e., physical dimensions and relative locations). This paper describes one part of a larger project whose goal is to obtain an integrated symbolic description of a scene from range data obtained from multiple viewpoints.

Image segmentation is a highly scene dependent and problem dependent decision making or pattern recognition process. Knowledge about the class of imagess to be processed and the tasks to be performed, plays an important role. Two approaches that explicitly incorporate such knowledge are advanced for the class of images containing polygonal shapes. They can be generalized to other shapes by change of pre-processing steps. Inference is both data driven and goal driven. It is guided by meta rules that are fired by the outputs of preprocessing. Effective suppression of noise is achieved. The methods illustrate the potential of AI techniques and tools for low-level image understanding tasks.

A vision guided robot work station is described, which is able to track and grasp objects moving on a conveyor belt. The vision algorithms are implemented on a ICOS 20000 Vision Computer. The Unimation Puma 560 robot running under VAL II offers the possibility of real-time path control. All movements are under guidance of the vision computer without any default path being programmed. Via a serial link, continuous updates for position and velocity are given, until the robot arm is correctly positioned above the object. A M68000 microprocessor based interface establishes the necessary protocols for robot - image computer communication. It also minimizes the length of the messages and takes time delays into account. The underlying vision algorithms are based on multi-resolution curvature measures for the object contours. These contours are first encoded with the "reduced generalized chain code". The vision system includes an automated modelling facility and preliminary algorithms for the generation of optimal recognition strategies. In order to accelerate recognition and localization, the concept of feature saliency was adopted.

This Paper Discusses A High-Level Knowledge-Based Vision System. We Have Implemented A Human Face Recognition System In Ops83, A Rule-Based Production System Language. The System Currently Is Capable Of Identifying The Components Of A Human Face From An Input Set Of Line Segments In A Manner Consistent With That Performed By Human Observers. The System Consists Of Independent Modules (Knowledge Sources), Each Of Which Is Capable Of Recognizing/Detecting A Particular Component Of The Face (E.G. Eye, Nose, Etc.) In An Image. Each Knowledge Source Posts, To A Section Of Memory, Information Regarding What Is Recognized And An Associated Confidence Level. This Global Data Area, The Blackboard, Allows The Posted Information To Be Available To All The Knowledge Sources. The Sharing Of Results Among The Knowledge Sources Provides Flexibility Such That The System Operates With Incomplete Image Information. The Extracted Results From The Knowledge Sources Are Evaluated And The Highest Confidence Face Is Presented.

Complex moments have been considered as useful features for object recognition in general. In this paper, the usefulness of applying the complex moment features into the tactile image for object recognition has been explored. Some complex moment invariants have been derived and implementation of those features has been conducted. With those moment invariants, we can elimiate the effect of lateral displacement and rotation from the tactile images. Through the generation of a decision tree and the utilization of the complex moment features, the shape of the objects from the tactile sensor can easily be recognized.

A process of slip between the robot gripper and the object being held was hypothesized. Using multiple modes of contact sensing, experiments were performed to detect evidence that would confirm the hypothesis. Some evidence was found. The experiments, results, and their potential significance are discussed.

In this paper we shall present new strategies for extracting tactile primitives: hardness, texture and geometric features. The novelty comes from the use of a finger shaped very much like the human finger, which incorporates a ferroelectric polymer-based tactile sensor. It is contrasted with the standard flat shaped tactile pads currently commercially available. In particular, the flatly shaped feelers are limited in their exploration of cavities. Hence our experiments will concentrate on examples of various cavities. We shall use vision for making hypotheses on the existence of the cavity and then use tactile exploration not only for verification but also for determination of the exact profile of the cavity. The latter is known to be very difficult to obtain solely by vision, since cavities, especially deep ones, usually do not reflect much light.

This paper presents a robot vision system which is capable of recognizing objects in a 3-D scene and interpreting their spatial relation even though some objects in the scene may be partially occluded by other objects. An algorithm is developed to transform the geometric information from the range data into an attributed hypergraph representation (AHR). A hypergraph monomorphism algorithm is then used to compare the AHR of objects in the scene with a set of complete AHR's of prototypes. The capability of identifying connected components and interpreting various types of edges in the 3-D scene enables us to distinguish objects which are partially blocking each other in the scene. Using structural information stored in the primitive area graph, a heuristic hypergraph monomorphism algorithm provides an effective way for recognizing, locating, and interpreting partially occluded objects in the range image.

In this paper An Acoustic Imaging Recognition System (AIRS) will be introduced which is installed on an Intelligent Robotic System and can recognize different type of Hand tools' by Dynamic pattern recognition. The dynamic pattern recognition is approached by look up table method in this case, the method can save a lot of calculation time and it is practicable. The Acoustic Imaging Recognition System (AIRS) is consist of four parts _ position control unit, pulse-echo signal processing unit, pattern recognition unit and main control unit. The position control of AIRS can rotate an angle of ±5 degree Horizental and Vertical seperately, the purpose of rotation is to find the maximum reflection intensity area, from the distance, angles and intensity of the target we can decide the characteristic of this target, of course all the decision is target, of course all the decision is processed by the main control unit. In Pulse-Echo Signal Process Unit, we utilize the correlation method, to overcome the limitation of short burst of ultrasonic, because the Correlation system can transmit large time bandwidth signals and obtain their resolution and increased intensity through pulse compression in the correlation receiver. The output of correlator is sampled and transfer into digital data by p law coding method, and this data together with delay time T, angle information eH, eV will be sent into main control unit for further analysis. The recognition process in this paper, we use dynamic look up table method, in this method at first we shall set up serval recognition pattern table and then the new pattern scanned by Transducer array will be devided into serval stages and compare with the sampling table. The comparison is implemented by dynamic programing and Markovian process. All the hardware control signals, such as optimum delay time for correlator receiver, horizental and vertical rotation angle for transducer plate, are controlled by the Main Control Unit, the Main Control Unit also handles the pattern recognition process. The distance from the target to the transducer plate is limitted by the power and beam angle of transducer elements, in this AIRS Models, we use a narrow beam transducer and it's input voltage is 50V p-p. A Robot equipped with AIRS can not only measure the distance from the target but also recognize a three dimensional image of target from the image lab of Robot memory. Indexitems, Accoustic System, Supersonic transducer, Dynamic programming, Look-up-table, Image process, pattern Recognition, Quad Tree, Quadappoach.

To avoid jamming during robot assisted assembly, the forces and moments induced by centering and alignment errors must be limited. A strategy is presented which incorporates feedback information from a vibratory interference sensor into the robot position control loop. A digital search algorithm is used to reduce the centering errors and and to assemble the mating parts.

This paper discusses the design of a contact sensor for use with the Utah-MIT dexterous hand [Jacobsen, et al. 1984]. The sensor utilizes an 8 x 8 array of capacitive cells. These results extend the work presented in Siegel, Garabieta, and Hollerbach [1985], Siegel [1986], and the earlier work of Boie [1984]. Before the sensor itself is discussed, a general outline for de-signing contact sensors is proposed. By evaluating the issues raised by this outline, the ultimate usefulness of the final device is more assured. The task outline is shown below: . Ascertain the quantity to be sensed. . Ascertain how this quantity can be sensed. . Determine where the sensor is to be used. . Determine what transduction processes can be applied. . Select the most appropriate transduction process. . Model the transduction process. . Determine how the sensor should be fabricated. . Build and test the sensor. It should be noted that designing tactile sensors is an exercise in evaluating engineering tradeoffs. Each of the above steps constrain the entire design process. The outcome of the latter steps will often force reevaluation of decisions made in earlier steps.

A tactile sensor was designed into the finger tips of a multifinger dextrous hand. The sensor features complete coverage of the cylindrical and hemispherical end portions of the finger. An 8X8 array is used based on standard capacitive sensing techniques. An analysis of appropriate sensor depth has been used to allow accurate localization of contacts, and reduced spatial aliasing. A comparison has been done between the cylindrical sensor and a simple two dimensional stress strain model. The effect of skin thickness and sensor depth on sensitivity are analyzed. Preliminary methods for deter-mining contact locatiot, total force, and tangential force with just normal deflection sensors have been implemented.

This paper describes a bin picking system being installed two characteristic algorithms; one automatically constructs two dimensional models of structural industrial parts and the other recognizes parts having cylindrical components as main ones, using their models. Models are constructed automatically as much as possible and are modified interactively, if necessary. Parts may be stacked randomly. The recognition algorithm first extracts the main components and next searches for other components, utilizing their models. When models are too simple to detect orientations of parts, each part is looked by the other camera after being picked out from a bin. Our system uses only one image for a scene of stacked parts. The three dimensional positions of parts are obtained using both the results of recognition and a touch detector equipped in the gripper of a robot. Experimental results are presented finally.

This paper describes recent research in the Aerospace Human Factors Research Division at NASA's Ames Research Center to develop a glove-like, control and data-recording device (DataGlove) that records and transmits to a host computerin real time, and at appropriate resolution, a numeric data-record of a user's hand/finger shape and dynamics. System configuration and performance specifications are detailed, and current research is discussed investigating its applications in operator control of dexterous robotic end-effectors and for use as a human factors research tool in evaluation of operator hand function requirements and performance in other specialized task environments.

A sensor-based intelligent robot may be used for moving objects, moving levels or knobs, assembly parts, driving screws, etc. However, for many task environments some constraints in task space generally exist. In these cases, neither pure position nor pure force control is appropriate. There is a need to develop a hybrid controller to accommodate various task environments. Previous development for a hybrid controller requires a definition of constraints prior to proceeding, i.e., the task environment must be predefined. It is still far removed from those desired by the manufacturing industry for full-fledged manufacturing implementations. We have developed a new approach for enhancing the current hybrid control techniques. The major contribution of this proposed hybrid position/force controller is its capability for accommodating the compliance selection vector both in position and force with the task environment adaptively. The derivation of an adaptive selection vector by the use of minimum energy and parameter identification algorithms is described in detail. The illustrative example and simulation results will be presented.

This paper describes SIERA (System for Implementing and Evaluating Robotic Algorithms), which has been developed at the Laboratory for Engineering Man/Machine Systems (LEMS) at Brown University. SIERA was created to satisfy the requirement for a multiprocessor-based development system flexible enough to be used for research into new robotic algorithms, especially those that utilize externally sensed information, such as vision and force. A multiprocessor architecture has been developed that incorporates a tightly coupled bus-based system for real-time servoing and a loosely coupled point-to-point network for less time-critical operations. SIERA is capable of controlling many types of commercially available robots since all input and output is done via general-purpose I/O boards. Suitably constructed robot interface boards are used to condition all feedback signals and to amplify the control outputs to the proper drive levels. We have constructed robot interface boards for the IBM 7565 and PUMA 560 manipulators in LEMS, and have controlled both robots using SIERA. The operating system used for SIERA has been designed to provide maximum flexibility in implementing new robotic algorithms. The concept of programming levels has been introduced to classify the different ways SIERA can be utilized�for simple robot control, for robotics research, and for system enhancements. The main benefit of SIERA is that it is now possible to experimentally implement and evaluate a variety of algorithms in areas such as compliant control, visual servoing, and inverse kinematics.

A practical learning control system is described which is applicable to complex robotic systems involving multiple feedback sensors and multiple command variables during both repetitive and nonrepetitive operations. The learning algorithm utilizes the Cerebellar Model Arithmetic Computer (CMAC) neural model developed by Albus. In the controller, the learning algorithm is used to learn to reproduce the nonlinear relationship between the sensor outputs and the system command variables over particular regions of the system state space. The learned information is then used to predict the command signals required to produce desired changes in the sensor outputs. The learning controller requires no a priori knowledge of the relationships between the sensor outputs and the command variables. The results of learning experiments using a General Electric P-5 manipulator interfaced to a VAX-11/730 computer are presented. These experiments involved learning to use video image feedback to track three dimensional task trajectories relative to objects moving on a conveyor. No a priori knowledge of the robot kinematics or of the conveyor speed or orientation relative to the robot was assumed. In all experiments, control system tracking error was found to converge after a few trials to within error limits defined by the resolution of the sensor feedback data.

Hypercube multiprocessors possess many attributes advantageous for autonomous, intelligent robots which operate in time-critical fashion in hostile environments. However, their message-passing architecture presents certain dif-ficulties in rapid changes of plans in response to unpredictable events in the environment. The status of a "virtual-time" operating-system shell, providing functions which facilitate such responses, is described.

Production lines with limited storage capacities can be modelled as cyclic queueing networks with finite buffers and general service times. A new technique, called perturba-tion analysis of discrete event dynamic systems, is applied to these queueing netowrks. Estimates of the gradient of the system throughput are obtained by perturbation analysis based on only one sample trajectory of the system. A modified Kiefer-Wolfowitz stochastic optimization procedure using the perturbation analysis estimates of gradients is proposed. This procedure possesses better convergence properties than the basic Kiefer-Wolfowitz procedure.

This short paper describes recent work at the IBM T. J. Watson Research Center directed at developing a highly flexible computational architecture for research on sensor-based programmable automation. The system described here has been designed with a focus on dynamic configurability, layered user inter-faces and incorporation of sensor-based real time operations into new commands. It is these features which distinguish it from earlier work. The system is cur-rently being implemented at IBM for research purposes and internal use and is an outgrowth of programmable automation research which has been ongoing since 1972 [e.g., 1, 2, 3, 4, 5, 6] .

Our experience indicates that it is possible to implement a certain degree of intelligence in a robot using a hierarchical control architecture. The structure is that of a community of sensor expert systems modules coordinated and controlled by a higher-level expert system. Research in the optical module is discussed with emphasis on an expert system autonomous calculation of object holdsites and manupulator grip-vectors. Design of the expert sysem was validated through simulated tests.

Automatic recognition, inspection, manipulation and assembly of objects will be a common denominator in most of tomorrow's highly automated factories. These tasks will be handled by intelligent computer controlled robots with multisensor capabilities which contribute to desired flexibility and adaptability. The control of a robot in such a multisensor environment becomes of crucial importance as the complexity of the problem grows exponentially with the number of sensors, tasks, commands and objects. In this paper we present an approach which uses CAD (Computer-Aided Design) based geometric and functional models of objects together with action oriented neuroschemas to recognize and manipulate objects by a robot in a multisensor environment. The hierarchical robot control system is being implemented on a BBN Butterfly multi processor. Index terms: CAD, Hierarchical Control, Hypothesis Generation and Verification, Parallel Processing, Schemas

This paper describes the architecture and Ada language implementation of a process-level run-time control subsystem for the Jet Propulsion Laboratory (JPL) telerobot system. The concept of run-time control in a combined robot-teleoperation environment is examined and the telerobot system at JPL is described. An Ada language implementation of the JPL Telerobot Run-Time Controller (RTC) is described by highlighting the functional behavior of the subsystem, defining the internal modules, and providing a functional flow time sequence of internal module activity.

The robot control C library (RCCL) is a high-level robot programming system developed at Purdue University as a result of research funded by the National Science Foundation and the French CNRS. RCCL enables a programmer to employ a set of system calls to specify robot manipulator tasks. The system provides flexible geometry kinematic transforms, hybrid force and position control, updatable world representation, functionally defined motions, and portability. RCCL allows for position matrix transform operations and the generation of robot trajectories in joint or Cartesian modes. It is intended to overcome the limitations of dedicated robot controller languages. For proprietary commericial robot languages like VAL, the fact that source code is not available rules out the possibility of modifications required to develop advanced control algorithms. In contrast, RCCL offers the major advantages of modifiable source code, sensor-oriented control, transportability between manipulator configurations and host computers, and computational speed necessary to execute advanced control algorithms necessary for real-time operation. In cooperation with JPL, RCA is completing RCCL implementation on a DEC microVAX-II as a high-level controller for the PUMA 762 robot, with minor changes to JPL's Berkeley 4.2 UNIX operating system. Using the same operating system, JPL plans to implement RCCL on a micro-VAX for their PUMA 560. These installations provide essentially unlimited memory, an efficient programming language for software development, and accessible source code that makes possible the real-time capability to meet the evolving needs of robot control technology. The Robotics Laboratory at McGill University provided consultation and direct support. The Laboratory successfully implemented RCCL on a VAX 11/750 for their PUMA 260 robot.

Development of programming language research in robotics and manufacturing has stressed the description of actions that constitute a robot task rather than the determination of actual robot movements. For example, the statement "turn the crank on the machine" could cause a task oriented system to automatically plan the necessary robot movements. We show that the latest stage of development in task languages is the inclusion of craft skills that achieve both brevity of expression and represent the sensory-kinematic skills that a human expert would have. A world class pianist knows how to strike the keys, not just that certain keys should be struck. A craft program to solve this kind of problem must be phrased as a multi-part dialog, because solutions must be tried, modified, and retried, interactively with the system components. The craft language (and dialogs) described in this paper is used to program a robotic workstation (i.e., robot, machining center, sensors, fixtures) so that parts can be accurately manufactured with a minimal amount of setup and programming. This work integrates research in task planning and expert systems with new ideas for acquiring sensory-kinematic skills from human experts.

The use of interactive computer graphics for simulation and off line programming provides a powerful tool in implementing robots. This capability essentially became available with the CAD/CAM system. This paper will present the theory of robot modeling and simulation techniques. An over-view of CAD/CAM system in robotic application, such as robot off-line programming, simulation and workcell layout will also be represented.

Most of today's industrial robots do highly repetitive tasks that require no human intervention for extended periods of time. It is, therefore, not too wasteful of operators' time when the destination of the end effector of such a robot must be modified occasionally by reprogramming the robot controller. In contrast, where daily tasks are varied and dependent on operator perception and judgement, robots have been excluded. We are investigating the use of pointing to specify "where" and in "what orientation" a robotic action is to be performed while voice or a keypad is used to determine "which" pre-programmed subroutine is to be executed by the robotic tool at the specified site. We are evaluating the relative advantages of voice and additional gesturesfor modifying "gesture-designated" end-effector position. We believe that the combination of gesture and voice for robot control will allow shop-floor personnel to efficiently and productively supervise multiple robotic tools work-ing on non-repetitive tasks that have previously been resistant to automation.

A survey of ten robot manufacturers has produced a task taxonomy of factory-floor programming and a compilation of hardware and software interface designs, the first undertaken in industrial robotics. This data base suggests the need for increased application of human factors engineering.

An integrated environment for the design and evaluation (from a human factors point-of-view) of human-machine interfaces is proposed. Four major components of an experimental environment currently under study are identified and discussed. A scenario highlighting the relationships of these four components in an integrated operational environment is presented. Current status, issues to be addressed and future plans for this activity are discussed.

A head-mounted, wide-angle, stereoscopic display system controlled by operator position, voice and gesture has been developed for use a multipurpose interface environment. The system provides a multisensory, interactive display environment in which a user can virtually explore a 360-degree synthesized or remotely sensed environment and can viscerally interact with its components. Primary applications of the system are in telerobotics, management of large-scale integrated information systems, and human factors research. System configuration, application scenarios, and research directions are described.

RCA's Advanced Technology Laboratories (ATL) has implemented an integrated system which permits control of high level tasks in a robotics environment through voice input in the form of natural language syntax. The paper to be presented will outline the architecture used to integrate voice recognition and synthesis hardware and natural language and intelligent reasoning software with a supervisory processor that controls robotic and vision operations in the robotic testbed. The application is intended to give the human operator of a Puma 782 industrial robot the ability to combine joystick teleoperation with voice input in order to provide a flexible man-machine interface in a hands-busy environment. The system is designed to give the operator a speech interface which is unobtrusive and undemanding in terms of predetermined syntax requirements. The voice recognizer accepts continuous speech and the natural language processor accepts full and partial sentence fragments and can perform a fair amount of disambiguation and context analysis. Output to the operator comes via the parallel channel of speech synthesis so that the operator does not have to consult the computer's CRT for messages. The messages are generated from the software and offer warnings about unacceptable situations, confirmations of actions completed, and feedback of system data.

Preliminary results of a system that uses model descriptions of objects to predict and match features derived from aerial images are presented. The system is organized into several phases: 1) processing of image scenes to obtain image primitives, 2) goal-oriented sorting of primitives into classes of related features, 3) prediction of the location of object model features in the image, and 4) matching image features to the model predicted features. The matching approach is centered upon a compatibility figure of merit between a set of image features and model features chosen to direct the search. The search process utilizes an iterative hypothesis generation and verication cycle. A "search matrix" is con-structed from image features and model features according to a first approximation of compatibility based upon orientation. Currently, linear features are used as primitives. Input to the matching algorithm is in the form of line segments extracted from an image scene via edge operatiors and a Hough transform technique for grouping. Additional processing is utilized to derive closed boundaries and complete edge descriptions. Line segments are then sorted into specific classes such that, on a higher level, a priori knowledge about a particular scene can be used to control the priority of line segments in the search process. Additional knowledge about the object model under consideration is utilized to construct the search matrix with the classes of line segments most likely containing the model description. It is shown that these techniques result in a, reduction in the size of the object recognition search space and hence in the time to locate the object in the image. The current system is implemented on a Symbolics LispTM machine. While experimentation continues, we have rewritten and tested the search process and several image processing functions for parallel implementation on a Connection Machine TM computer. It is shown that several orders of magnitude faster processing rates are achieved, as well as the possibility of entirely new processing schemes which take advantage of the unique Connection Machine architecture.

The Robotic Locating System (RLS) is a multisensor laboratory test bed designed to evaluate the feasibility of measuring target range and orientation in various robotic environments. The targets are considered to be cooperative and typically have LEDs or reflective material placed on them. Laser illumination in combination with video sensing is the primary sensor system. A pair of optical position sensors and an array of ultrasonic ranging sensors to augment system performance. This paper will report on the sensitivity study being done to determine the resolution and accuracy of the measurement process. In particular, details of measuring all six degrees of freedom (DOF) of a rectangular target with four LEDs or four spots of reflective tape will be presented.

We describe some experiments in real-time 3-D object classification using a learning system derived from a general neural model for supervised learning. The primary advantages of the learning system are its ability to learn from experience to recognize patterns and its inherent massive parallelism. Our motivation is to examine the feasibility and merits of the learning system in a simple machine vision problem.

Current algorithms used in image processing and image understanding applications impose diverse and demanding computational requirements on processor architectures. While general purpose computers can provide the needed types of computation, the performance requirements driven by high-rate imaging sensors force system architects to use highly parallel and/or specialized hardware to attain an efficient system realization. To date, end-to-end systems capable of attaining real-time performance at full sensor data rates have relied on a mixture of processor types generally including specialized hardware for low-level pixel processing and general purpose processors for abstract symbolic data manipulation. Such systems are plagued by several problems: a clean division of algorithms among processor types is difficult and can impose artificial constraints; each processor type requires a different software environment making a unified programming methodology difficult; specialized pixel processing hardware frequently possesses little or no programmability; and programmable processors rely on microcoding to maximize concurrency but make the conversion from course-grained high-order languages difficult. Honeywell's Macro architecture uses data-driven data flow techniques developed for supercomputing to attain high performance parallel computation while simplifying high-order language programming. A compact and efficient processor realization is attained through a system organization which takes advantage of the macro nature of signal and image data and through hardware constructs which make low-level pipelining invisible to the user.

Conventional robotic welding systems can only be used in applications where parts are highly repeatable and well fixtured. In this paper, vision sensing and processing techniques that permit in-process determination of the position and detailed three-dimensional surface geometry of a weld joint are presented. Structured lighting in the form of a cone of laser light and specialized vision processing schemes are used to obtain three-dimensional geometric surface descriptions from a two-dimensional TV image. This geometric description is used to define joint position, surface orientations, and a variety of cross-sectional measurements such as fill volume, preparation angles, gap size, and presence and dimensions of a previously deposited weld bead or tack weld. The visual feedback is then used for real time control of the torch position relative to the weld joint, and for in-process adjustment of welding process parameters such as welding arc voltage, and wire feed rates.

Today, in industrial environment, there is an increasing requirement of vision systems. Of course, some big systems exist, but they are often too expensive in relation to the aimed applications. The described system forms a low cost and well fitted solution, with great adaptability. This system has a modular architecture. The basic elements are image digitization module, specific processor module, microprocessor-based module. The system architecture can be easily redefined for each application by judiciously assembling the basic modules. The whole system appears like a net of interconnected processors, exchanging informations. The internal organization gives the hierarchy between processors and defines "master" and "slave" processors. Interconnection and number of processors depend on the application, expected preformances, computation complexity, number of cameras, and so on. The system modularity makes it very adaptative to many applications. Its realization with recent technology and popular microprocessors breaks costs down.

This paper presents the hardware and software architecture developed to control the Utah-M1T dexterous hand. This work is part of a joint project between the Center for Engineering Design at the University of Utah, and the MIT Artificial Intelligence Laboratory.