A hierarchial multi-level feature-space pattern recognition system is described. Multi-class distortion-invariant object identification is the purpose of this study. Attention is given to dimensionality reduction (to simplify computations) and to the use of non-unitary transformations (to achieve discrimination). A Fourier transform feature space is used. However, our basic hierarchial concepts, our theoretical analysis, and our general conclusions are applicable to other feature spaces. The use of intensity versus phase features is studied and the performance of our system in the presence of noise is studied. Quantitative experimental data on 2 two-class pattern recognition databases are provided.

An approach to pattern recognition applying computer vision system after the method of characteristic parameters such as perimeter, area, number of holes, maximum and minimum moment of inertia, maximum, minimum and mean radius and vector of holes is presented. All such parameters are computed by the aid of preselected characteristic contour points. The position and orientation of objects are also determined. Nearest neighbour method is applied for identification.

Two conceptually new algorithms are presented for segmenting textured images into regions in each of which the data is modelled as one of C non-causal 2-D Markovian Stochastic Processes. The algorithms are designed to operate in real time when implemented on new parallel computer architectures. A doubly stochastic representation is used in image modelling. Here, an auto-normal (Gaussian) process is used to model textures in visible light and infrared images, and an auto-binary field is used to model apriori information about local image geometry. Image segmentation is realized as true maximum likelihood estimation. The first segmentation algorithm is hierarchical and uses a pyramid-like structure in new ways that exploit the mutual dependencies among disjoint pieces of a textured region. The second segmentation algorithm is a relaxation-type algorithm that arises naturally within the context of these non-causal Markovian Processes. It is a simple, true maximum likelihood estimator. The algorithms can be used separately or together. These issues and subtleties concerning the use of the Markovian processes are discussed.

Surface finishing operations carried out on castings are difficult to automate. This is partly due to the irregular and random nature of defects. Optimal application of robot finishing techniques therefore requires in-process surface condition information in order to implement an adaptive control strategy. This paper describes a robotic device which incorporates a model based machine vision inspection system to identify and to locate defective surface areas. Viewing arrangements are preplanned from the component surface model and a knowledge of the robot kinematics. The model assists in the interpretation of image data which provides information for the overall adaptive control of the abrasion process.

The i-bot system is a flexible, adaptive and programmable sensory device for industrial robots. The system consists of independent modules for vision and gripper control, that guide the robot to perform acquisition of jumbled or isolated parts from a bin. The acquired parts are transported to a fixture and properly aligned by the robot-arm. The parts in the bin are identical except for minor dimensional and textural variations. These parts may be colored, contain graphics, possess multiple diameters or other cross sections, mixed material content and reflectivities. The system operates by taking a high resolution grey scale image of the bin of parts using an overhead camera under ambient illumination. The image is preprocessed by video range expansion, windowing, histogram based enhancement techniques and is segmented into multiple clusters by analyzing spatial characteristics. Algorithms are developed for fast and iterative data compression and for estimation of location and orientation of a cluster irrespective of its shape. The photobeam, collision, and pressure sensors on the parallel jaw gripper aremonitored during the acquisition process. Experiments have been performed with fuel filters, curved plier blanks, multicolored markers, "L" shaped fuel links, ball bearings and a mixture of semi-rusted threaded and unthreaded bolts of various lengths. In one image processing cycle of 2 seconds, i-bot can compute locations and orientations of a maximum of 3 work-pieces at a successful acquisition rate of approximately 95%.

For many applications, eye-in-hand robot vision has several advantages over the use of a fixed external camera. Even smaller solid state cameras are too bulky for gripper mounting and, therefore, limit the gripper's movement for many operations. A novel approach for eye-in-hand robot vision is proposed utilizing, a) coherent fiber-optic bundles for carrying light from an object to be imaged, and b) photodiodes to convert this light into electrical signals for processing. The feasibility of this concept has been experimentally demonstrated by developing a fiber-optic eye-in-hand system, mounting the scanner head on a gripper and obtaining good images under back and front light conditions.

An original imaging system using eddy current sensors is being presented. The linear disposition used gives the picture of the profile of a metallic surface in front of the sensor , the main application being weld guidance systems. Each sensor in this linear structure gives a measure of the distance of the metallic surface in front. Individual sensors and signal acquisition use an original design intended to achieve maximum directivity and dynamic range, and insensitivity to the conductivity of the metal and to electromagnetic interferences. Simultaneous, in phase, excitation of all the sensors turns out to give an improved spatial resolution. The signals from individual sensors are multiplexed, digitalized and used to address a preregistred linearisation memory in order to obtain a first raw image of the profile. To improve this image a digital treatment is used which achieves non linear deconvolution, and interpolation between the observation points.

For robot ano automated arc welding a compact 3D vision sensor has been developed. It generates data to correct the preprogrammed welding path. It can find the starting point of toe seam, calculate the seam volume for process control and detect tack welds, all in real-time with submillimeter resolution. Mounted to the robot hand the He-Ne laser-based scanning triangulation sensor takes distance measurements to the workpiece at a 2000/s rate. Preproduction models were tested in operation for the effects of arclight, heat, smoke, spatter ano e.m.i., using various combinations of the welding_ parameters: process (MIG, TIG), mode (short cktg, globular, spray), welding current and voltage, protective gas (He, Ar, CO2) and travel speed. Included in the paper is a discussion of environmental influences and of closed-loop operation.

A dramatic increase in the use of industrial robots is expected in the United States within this decade. This increase increase will require robot vision systems so that the machines can respond and adapt to their environment, i.e., be intelligent. These advances can lead to new methods for use in flexible manufacturing systems and automated factories. The purpose of this paper is to describe a technique for three dimensional robot vision using a single projected image. The method involves automatic location of projected vertex points to determine the three dimensional surface characteristics of an object. The method is robust enough to permit measurement of featureless curved objects. Examples are given which demonstrate the simplicity of the technique. The significance of this method is that it is simple enough for use in robot control or inspection applications.

Absolute location of a mobile robot is necessary to improve the autonomy of vehicle built for hostile environments. We are developing a scanning laser range finder based on triangulation to get range data about the edges of a cylindrical polyhedral world. From the matching between the measurements and data computed from a model of the a priori known environment, the position of the robot is deduced accurately.

An electro-optical subsystem which adds the X-Y random access motion and focus of human vision to robotics vision systems is described. The capabilities of the human eye and that of the new ocular manipulator to perform industrial inspection and solve character recognition problems are compared. Both the optical servo and computer control aspects of the subsystem are discussed. Applications to specific industrial problems requiring optical image manipulation are also described.

A system of three dimensional vision and space perception for industrial robots is described. The system is based on binocular or two camera stereopsis and is derived from earlier successful systems used in the automation of map compilation from overlapping aerial photographs. It is suitable for the identification and inspection of irregularly shaped objects and for motion control as in orientation, shaping, fabrication and assembly operations and for the guidance of automated vehicles in active areas.

This paper describes decomposition of an object into upright hexagonal cylinders to represent the space occupied by the object. The representation results in a 7-order tree data structure, called septree. A hierarchical addressing system, the Generalized Balanced Ternary (GBT), is employed as an addressing mechanism for the cylinders. The objective of this paper is to present efficient computational algorithms for manipulating the tree representation as the object undergoes translation and limited rotation through space.

The complete information about contour structures of a gray level image can be encoded by a hierarchical procedure. Continuity of lines and edges is explicitly checked and extended continuous structures are encoded at high levels of a hierarchical contour code (HCC), while less extended details are available at lower levels. Though contour information is processed in seperate spatial frequency channels, correspondence is arranged by the special structure of the code. By this approach, a pattern recognition system may be split into a general purpose front end, encoding the complete contour information of a gray level image, and into a recognition part with top down access to this code, using only as much information as necessary for an unambiguous identification of the scene. This structure is very promising for robot vision applications, as the general purpose coding system can be implemented in real time hardware and as the recognition part is relieved of processing the complete gray level image.

This paper examines the application of the "shape from shading" technique of surface measurement for depth perception and object recognition. An examination of several models for surface shading prediction as applied to scene synthesis is presented. Also discussed is the reverse process of using these surface shading models for image analysis for depth perception. The surface measurements are then used to fit simple surfaces to objects for recognition.

While performance and working dexterity of simulation-systems and software-tools for computer aided manufacturing have lead to a rapid increase in the numbers of succesfull implementations, there is still a lack of knowledge concerning systematic approaches for development and use of complex, sensor-controlled equipment in flexible automation. This is because of two facts; on the one side sensor systems in the form of electro-optical devices and picture processing units as flexible solutions for industrial applications have extremly different capabilities and it is a hard task to decide what system to use under economic considerations and on the other hand we often find a rather simple straight forward approach for the sensor problem appearing in a specific production task. This usually leads to an one-pur-pose sensor system nonconformant with the flexible basis equipment, or even to a solution without any sensors but with additional non-flexible peripheral hardware. When few years ago electro-optical sensors were developed for assisting industrial robots in parts-handling, there were problems in justifying the overall enconomics of such systems. This has lead to them being used in Quality-control, where especially cost inten sive high-precision sensors were absolutely necessary. However during this period rapid progress has been made increasing the performance and reliability of such systems resulting in a migration back to the robotic field. Based on a few typical examples we will discuss various aspects of the implementation of sensor devices with varying decrees of complexity and costs, both in the fields of flexible production and quality control.

A fast stereo-matching algorithm designed to operate in the presence of noise is described. The algorithm has its roots in the zero-crossing theory of Marr and Poggio but does not explicitly match zero-crossing contours. While these contours are for the most part stably tied to fixed surface locations, some fraction is always perturbed significantly by system noise. Zero-crossing contour based matching algorithms tend to I- very sensitive to these local distortions and ar, prevented from operating well on signals with moderate noise levels even though a substantial amount of information may still be present. The dual representationâ€â€?regions of constant sign in the V2G convolution persist much further into the noise than does the local geometry of the zero-crossing contours that delimit them. The PRISM system was designed to test this approach. The initial design task of the implementation has been to rapidly detect obstacles in a robotics work space and determine their rough extents and heights. In this case speed and reliability are important but precision is less critical. The system uses a pair of inexpensive vidicon cameras mounted above the workspace of a PUMA robot manipulator. The digitized video signals are fed to a high speed digital convolver that applies a 32² VG operator to the images at a 106 pixel per second rate. Matching is accomplished in software on a lisp machine with individual near/far tests taking less than i3luth of a second. A 36 by 26 matrix of absolute height measurements - in mm - over a 100 pixel disparity range is produced in 30 seconds from image acquisition to final output. Three scales of resolution are used in a coarse guides fine search. Acknowledgment: This report describes research done at the Artificial Intelligence Laboratory of the Massachusetts Institute of 'Technology Support for the laboratory's artificial intelligence research is provided in part by the Advanced Research Projects Agency of the Department of Defense under Office of Naval Research contract N00014-80-C-0505 and in part by National Science Foundation Grant 79-23110MCS.

The purpose of this paper is to describe a project which makes three dimensional measurements of an object using a robot arm. A program was written to determine the X-Y-Z coordinates of the end point of a Minimover-5 robot arm which was interfaced to a TRS-80 Model III microcomputer. This program was used in conjunction with computer graphics subroutines that draw a projected three dimensional object.. The robot arm was direc-ted to touch points on an object and then lines were drawn on the screen of the microcomputer between consecutive points as they were entered. A representation of the entire object is in this way constructed on the screen. The three dimensional graphics subroutines have the ability to rotate the projected object about any of the three axes, and to scale the object to any size. This project has applications in the computer-aided design and manufacturing fields because it can accurately measure the features of an irregularly shaped object.

The addition of vision or three-dimensional imagery to a robot significantly enhances its capabilities to perform not only well-defined tasks, but new tasks as well -- tasks that could not be done by a "blind" robot. At the next level, the robot would have a capability of acquiring, processing and analyzing a three-dimensional image of a "real" scene. This paper presents several alternate ways that such three-dimensional imagery can be acquired and represented. Some emphasis will be placed on one approach -- the use of laser radar to obtain range data. However, the primary objective in this paper is to identify and discuss the major problems associated with the acquisition and effective use of this type of imagery, especially in robot vision applications.

An increasingly common technique for extracting three-dimensional (3-D) information for robotics and automation involves a camera and laser spot projector. This paper presents an analysis of the accuracy of the 3-D measurement using this approach. The sources of error are assumed to arise from limited precision of the camera and imperfect system calibration. The resultant error function relates the system accuracy to both internal camera and external system geometry. For a typical system configuration, utilizing solid state cameras, data show that internal camera parameters have negligible effect on the system accuracy. The greatest determinant of system performance is the system alignment. This is followed closely by the camera orientation angle.

An approach to 3-D recognition of randomly oriented parts with known geometry is described. The method relies on line and corner data extracted in a preprocessing stage. A single 3-D geometrical model of the part is used. The part description is CAD-like. An Augmented Transition Network (ATN) controls an analysis-by-synthesis search procedure that is based on hypothesis generation and verification. A state-space search technique is used.

Sensing is the mechanism by which robots receive information from the world around them. Robot sensing, and in particular, robot vision is the basis for providing sensory inputs to a smart robot. Devices that make robot vision possible are typically classified as electro-optical sensors. Thus a critical review of robot vision is essentially a review of current and predicted capabilities for imaging electro-optical sensors. The critical element in an electro-optical sensor is the detector and most detectors are semiconductor devices. Thus we conclude that robot vision is closely tied to the state of the art in the development of photo-sensitive semiconductor materials and in the fabrication of large scale arrays of detector elements. This paper begins with a review of the fundamental principles that are important to electro-optical sensors. Selected state of the art sensor systems are next analyzed to show expected performance. Finally a review is given of detector technology as related to robot vision including a comparison of materials, array sizes, performance and cost.

A series of experiments will be described involving: (a) electric light bulbs, (b) glass vials, (c) eggs, (d) boxes or bags. In each case, consideration will be given to the visual, rather than the manipulator aspects of palletising/packing these objects. Emphasis will be placed on the optical/ illumination techniques and the image processing needed to identify the significant features in the picture data.

Early stages of visuo-motor interaction are considered with regard to dynamic scene analysis. Target fixation and tracking is distinguished from target visual analysis. The notion of target specification is elaborated upon. The use of ternary dynamic images is shown as an example of target tracking.

A method for segmentation of images is introduced, and a technique for detecting ellipses in the regions is discussed. The former is based on iteratively dividing a histogram into some modes in taking account of a measure for adjacent separability. The latter consists of classifying behavior of parameters in a quadratic equation obtained by using a recursive least-squares estimator. Some experimental results for applying these methods to machine parts are shown.

The tactile sensor described in this paper employs two stages of transduction in converting impressions on its touch surface into conveniently processed electrical signals. One stage is a pure mechanical deflectometer while the other is an arrangement using electro-optical transducers. This transduction concept contributes to a robust sensor design suitable for operation on the factory floor. Design, performance, and application of the tactile sensing system are described. The application is an operation involving assembly of a thin flexible diaphragm in the lip of a rigid cap.

A retractable fiber optic sensor has been installed to control a robot for the task of cleaning metal castings with a plasma-arc torch. A number of other sensor types were considered, and rejected, for this application. A single-pass search-and-cut method was successfully demonstrated, using a specially designed control interface.

The application of color television technology to computer color perception and image processing is demonstrated using photographic color control patches and gray scales.

A fish-eye lens image can include information covering a 2Tr steradian field of view. Such a lens can be used as an element of the vision system of a robot. These images, which are recorded digitally, appear with an unavoidable inherent distortion which can be corrected with image processing techniques. A method for geometric restoration of such distorted images is derived and its resolution and error are analyzed. A system which may serve as a robot's vision system is described, along with examples of "distorted" and "undistorted" images to clarify the results of the above method.

This paper describes a contour tracking algorithm for images which have been hexagonally sampled. Both formal and informal descriptions of the procedure as well as performance statistics are included.

Quality control is a problem in the whole area of industrial production. In-line quality control increases the productivity by enabling greater velocity of the production processes as well as by assuring higher standards of the products. Possibilities of the application of electro-optical systems to in-line quality control are discussed. As an example the in-line inspection of a gasket attached to a water pump is described. Additionaly it is shown how electro-optical systems can support robots to correct parts-position and orientation according to the assembly task.

The outline corner filter is a very fast 2-D outline data compression algorithm. Input to the algorithm is a chain-coded outline, or connected sequence of points; output are the co-ordinates of the vertices of an approximating polygon. The style and accuracy of the approximation are controlled by parameters, and are widely variable. The accuracy is contained within defined limits, which is important for robot grasping. The algorithm has been carefully optimised to run very rapidly in a cheap microprocessor. To that end, it uses no multiplication or division, and contains no iterated loops.

In this paper results of researches which go in favour of association of abstract spaces with the natural one are presented.Results were obtained applying approach of symmetry theory and action principle.The exi stance of a specific space,which is connecting the above ones and called "linking" space is theoretically described. The possibility for defining and realization of "linking" space within the computer memory,with all peculiarities it must posses is shown.An example for using the linking space for the visual space perception is given. A short description and conclusions make the end of the paper.

A typical computer vision system has an analog (i.e. TV camera) front end that delivers a two-dimensional image to digital hardware. Often, the initial computations consist of extraction of features from this image. We propose here an analog preprocessor that can directly deliver a variety of features or perform certain kinds of computations on the input image in a time comparable to the scan times of most TV systems. The system operates directly on projections (i.e. line integrals) of the two-dimensional image rather than the image itself. Two reasons for this are: (1) certain operations are performed more naturally in the projection space, (2) the projections can be processed as one-dimensional temporal signals; devices for these operations have significant advantages over two-dimensional spatial signal processors. We show implementations for performing the following operations: (1) features consisting of the image,Fourier power spectrum integrated along radial pie-wedge segments and concentric arcs. (2) Computation of the Hough transform for detecting line segments. (3) Computations of moments invariant to translation, rotation, and geometric scaling. (4) Computation of certain geometric features with similar invariance properties.

Through the use of intelligent sensors such as vision, conventional machines are able to react and decide for themselves how to adapt to changing circumstances. This paper reviews the increased scope that can be achieved in manufacturing processes by automatically controlling groups of machines under the master control of intelligent sensors. The paper will refer to examples from a wide range of industries.

The automation of car body spray booth operations employing paint spraying robots generally requires the robots to execute one of a number of defined routines according to the car body type. A vision system is described which identifies a car body type by its shape and provides an identity code to the robot controller thus enabling the correct routine to be executed. The vision system consists of a low cost linescan camera, a flucrescens light source and a microprocessor image analyser and is an example of a cost effective, reliable, industrially engineered robot vision system for a demanding production environment. Extension of the system with additional cameras will increase the application to the other automatic operations on a car assembly line where it becomes essential to reliably differentiate between up to 40 vatiations of body types.

Unlike NC machine-tools, where the tool path is primary and the product shape results from it, in arc welding the product is primary, and the welder - human or robot - has to contend with poor fitups, bad preparations, inexact positionning etc. All this means one thing - adaptivity. The axtent to which this is reasonable is discussed, and then a research project, conducted at IICR with the aim to create an adaptive GMA robot, is presented.

The present paper discusses the design of the control system for an adaptive arc-welding robot. The upper limits of intelligence of these robots are discussed. Based on this, a decomposition of the task for controlling the robot is realised. The structure of the diffe-rent levels and communication between them are explained. The basic program modules at all levels as well as the vertical and horisontal data flows are analised. The teaching process is presented as a process of model construction, describing the functioning of the separate system units.

This paper describes the mechanical behaviour of a flexible, cylindrical peg during it's insertion. Three cases are examined according to the displacement of the contact points. In each case, a jamming criterion is given and the state of the peg is determined using only the informations of the positionning system and the force sensor of the robot. The control law for the insertion is deduced from this analysis.

The use of industrial robots is a well established method for increasing flexibility and the degree of automization in manufacturing industry. Nevertheless, the ilsage of robots is often limited by the need for special peripheral equipment. Conventional sensors enable a robot to "feel" the workpiece and tools, but it is only first when the robot is provided with vision that it can identify and localize objects. The described robot vision system is completely integrated with the robot control system, thus resulting in lower costs for engineering design, installation,commissioning,operation and maintenance while increasing the system's reliability. It uses the same method for man-machine communication as the robot system does and can be utilized by the workshop personnel. It is adapted to industrial lighting intensity, which means that special lighting arrangements are not required.

On-line monitoring of the cutting process in CNC lathe is desirable to ensure unattended fault-free operation in an automated environment. The state of the cutting tool is one of the most important parameters which characterises the cutting process. Direct monitoring of the cutting tool or workpiece is not feasible during machining. However several variables related to the state of the tool can be measured on-line. A novel monitoring technique is presented which uses cutting torque as the variable for on-line monitoring. A classifier is designed on the basis of the empirical relationship between cutting torque and flank wear. The empirical model required by the on-line classifier is established during an automated training cycle using machine vision for off-line direct inspection of the tool.

We introduce a new proximity sensor and show its advantages for robotic assembly and manipulation of recessed parts. The sensor is based on the Reentrant-loop magnetic effect. We demonstrate experimentally that a simple embodiment of this sensor in a robot hand can detect position and orientation of marked objects within ri5 cm range, independently of the speed of approach.

An ideal integrated robotic welding system should incorporate off-line programmation with the possibility of real time modifications of a given welding programme. Off-line programmation makes possible the optimization of the various sequences of a programme by simulation and therefore promotes increased welding station duty cycle. Real time modifications of a given programme, generated either by an off-line programmation scheme or by a learn mode on a first piece of a series, are essential because on many occasions, the cumulative dimensional tolerances and the distorsions associated with the process, build up a misfit beetween the programmed welding path and the real joint to be welded, to the extent that welding defects occur.

The rapid pre-processor initially developed by NRCC and Leigh Instruments Inc. as part of the visual aid system of the space shuttle arm 1 has been adapted to perform real time seam tracking of multipass butt weld and other adaptive welding functions. The weld preparation profile is first enhanced by a projected laser target formed by a line and dots. A standard TV camera is used to observe the target image at an angle. Displacement and distorsion of the target image on a monitor are simple functions of the preparation surface distance and shape respectively. Using the video signal, the pre-processor computes in real time the area and first moments of the white level figure contained within four independent rectangular windows in the field of view of the camera. The shape, size, and position of each window can be changed dynamically for each successive image at the standard 30 images/sec rate, in order to track some target image singularities. Visual sensing and welding are done simultaneously. As an example, it is shown that thin sheet metal welding can be automated using a single window for seam tracking, gap width measurement and torch height estimation. Using a second window, measurement of sheet misalignment and their orientation in space were also achieved. The system can be used at welding speed of up to 1 m/min. Simplicity, speed and effectiveness are the main advantages of this system.

As the complexity of sub-sea maintenance and inspection tasks increases, and the search for further reserves of offshore oil goes into deeper and more hazardous environmental conditions, the need for autonomous working submersibles is increased. The total free swimming submersible system requires a scope that embraces the guidance and control of the vehicle and associated payload, the acquisition and processing of information, and the communication of that information between the vehicle and the surface operator. This paper describes an important aspect of the total system, namely, the technique employed in reducing the required bandwidth for the transmission of television pictures, through an acoustic communication link. In particular, the estimation of motion within an image sequence has been shown to offer a further refinement in bandwidth reduction when used in conjunction with interframe hybrid coding. Many competing techniques have been proposed in the literature. The objective of this study has been to compare a range of fundamental techniques, in order to optimise the communication system design for this particular application. An essential by-product of this investigation has been the development of a general image processing system that incorporates a set of interactive image processing modules and provides a real-time colour picture capture and display facility.

The Robotics Design and Construction Class at Sehome High School was developed to offer gifted and/or highly motivated students an in-depth introduction to a modern engineering topic. The course includes instruction in basic electronics, digital and radio electronics, construction skills, robotics literacy, construction of the HERO 1 Heathkit Robot, computer/ robot programming, and voice synthesis. A key element which leads to the success of the course is the involvement of various community assets including manpower and financial assistance. The instructors included a physics/electronics teacher, a computer science teacher, two retired engineers, and an electronics technician.