This paper presents a new approach to the transfer of technology in Image Understanding (RI). This approach is based upon the use of a shared software development environment for JU, called the KBVision System, which can serve to ease the transfer of computer vision algorithms. In particular, we discuss a DARPA-sponsored SBIR project in which the KBVision' System has been used to integrate and distribute algorithms from two very different research environments.

This paper describes a new method for building object models for the purpose of overlapped object recognition. The method relies on local fragments of the boundary to derive a set of autoregressive parameters that serve to detect similar boundary fragments. First a rule based algorithm which detects the occlusion of two or more objects is introduced. This algorithm makes use of aheuristic rule which take into account the number of intersection points of the boundary with a standard invariant shape and of global features (area, perimeter) to confirm the presence of occlusion. The object is then decomposed into visible parts by using first a polygonal approximation method and then the concave vertices obtained at the latter step. The decomposition algorithm prepares the input data for the description of the model and the object through the autoregressive filter method.

Recursion and feedback are two important processes in image processing. Image algebra, a unified algebraic structure developed for use in image processing and image analysis, provides a common mathematical environment for expressing image processing transforms. It is only recently that image algebra has been extended to include recursive operations [1]. Recently image algebra was shown to incorporate neural nets [2], including a new type of neural net, the morphological neural net [3]. This paper presents the relationship of the recursive image algebra to the field of fractions of the ring of matrices, and gives the two dimensional moving average filter as an example. Also, the popular multilayer perceptron with back propagation and a morphology neural network with learning rule are presented in image algebra notation. These examples show that image algebra can express these important feedback concepts in a succinct way.

This paper describes a new method for building object models for the purpose of overlapped object recognition. The method relies on local fragments of the boundary to derive a set of autoregressive parameters that serve to detect similar boundary fragments. First a rule based algorithm which detects the occlusion of two or more objects is introduced. This algorithm makes use of aheuristic rule which take into account the number of intersection points of the boundary with a standard invariant shape and of global features (area, perimeter) to confirm the presence of occlusion. The object is then decomposed into visible parts by using first a polygonal approximation method and then the concave vertices obtained at the latter step. The decomposition algorithm prepares the input data for the description of the model and the object through the autoregressive filter method.

The analytical and processing developments in the field of Image Understanding over the last 15 years have led to the creation of a set of processing tools for the detection, characterization (feature extraction), and classification of 2 dimensional signals. This set of tools is applicable to 2 dimensional signals other than the traditional "image" type signals. In particular, for passive sonar detection processing several 2 dimensional signal transforms are generated from the 1 dimensional sensor time series data. These transforms are selected in order to concentrate signal energy locally within the 2 dimensional transform. A classic example is the Lofargram which is a grequency versus time transform of the time series data. If the acoutic source is emitting tones (for example from machinery) then the Lofargram will contain line like structures.

As part of on-going studies of automated techniques for object recognition in imagery, recent experiments in two and three dimensions have produced promising results. Newly developed methods that exploit projectively invariant relationships in imagery are able to recognize the same object in images that differ in tilt, scale and rotation. Automatically extracted corner points are used as the base features in simple two-dimensional objects, and patches of known gray-value are used in threedimensional terrain perspective views. In both cases, projective invariants are calculated and compared with a catalog of archetypal values, resulting in successful identification of the objects within experimental error.

The Geometric Hashing technique developed by the NYU Courant Institute has been applied to various automatic target recognition applications. In particular, I-MATH has extended the hashing algorithm to perform automatic target recognition ofsynthetic aperture radar (SAR) imagery. For this application, the hashing is performed upon the geometric locations of dominant scatterers. In addition to being a robust model-based matching algorithm -- invariant under translation, scale, and 3D rotations of the target -- hashing is of particular utility because it can still perform effective matching when the target is partially obscured. Moreover, hashing is very amenable to a SIMD parallel processing architecture, and thus potentially realtime implementable.

We have developed a structured light system for noncontact 3-D measurement of human body surface areas and volumes. We illustrate the image processing steps and algorithms used to recover range data from a single camera image, reconstruct a complete surface from one or more sets of range data, and measure areas and volumes. The development of a working system required the solution to a number of practical problems in image processing and grid labeling (the stereo correspondence problem for structured light). In many instances we found that the standard cookbook techniques for image processing failed. This was due in part to the domain (human body), the restrictive assumptions of the models underlying the cookbook techniques, and the inability to consistently predict the outcome of the image processing operations. In this paper, we will discuss some of our successes and failures in two key steps in acquiring range data using structured light: First, the problem of detecting intersections in the structured light grid, and secondly, the problem of establishing correspondence between projected and detected intersections. We will outline the problems and solutions we have arrived at after several years of trial and error. We can now measure range data with an r.m.s. relative error of 0.3% and measure areas on the human body surface within 3% and volumes within 10%. We have found that the solution to building a working vision system requires the right combination of theory and experimental verification.

The success of machine vision systems in solving real-world problems will depend on how well they can balance the conflicting requirements of high accuracy, flexibility to operate under a wide range of environmental conditions, fast response time, and size constraints. Most machine vision systems developed in the past have sacrificed one or more of these factors in favor of the others. In this paper, we discuss our experience in developing high speed machine vision and world modeling systems for mobile robotics applications. A pipeline binocular stereo range detection system developed in our laboratory matches 256 x 256 pixel stereo image pairs in one second and generates 2- D and 3-D obstacle maps in near real-time. These obstacle maps then get integrated into pixel- and voxel-based dynamic world models. Using data provided by stereo cameras mounted on top of an indoor mobile robot, these systems have the capability to create very realistic models of the environment. An autonomous navigation system uses these environment models to successfully navigate a mobile robot in an indoor environment cluttered with dynamic and static obstacles.

Conventional fieldable signal processing systems utilize custom hardware manufactured and configured specifically for a single signal processing application. Developing new systems or reconfiguring existing systems involves great expense and time expenditure. We at Alliant Techsystems have developed a signal processing system based on commercially available hardware which is completely software programmable and yet small and fast enough to be used in fieldable multisensor signal processing applications. This paper will discuss Alliant's reconfigurable signal processing system.

Litton Data Systems has developed a hybrid ATR system using DataCube signal processing and imaging boards controlled by a host Sun workstation. The input image preprocessor includes detection algorithms that locate candidate targets in a simulated or real JR scene and segmentation algorithms that generate target outlines which are normalized for recognition processing by Litton's LIGHTMOD optical correlator. Target detection performs a raster scan over the image using nested boxes to find detection pixels based on intesity/variance contrasts. Segmentation operates on windows centered at the centroids of connected components of detection pixels. Optimal thresholds are determined from gray scale histograms. Multimodal optimization seeks to maximize edge/boundary proximities for candidate target and island components obtained by thresholding with respect to the optimal thresholds. Relaxation smooths the boundaries of optimal target components or merged target/island components. Normalization algorithms scale and center the component boundaries in a 128x128 LIGHT-MOD window. The LIGHT_MOD correlates the FT of the normalized outline with the FTs of a reference library of target outlines. Template matching is the basic technique used for recognition.

Applying distributed memory parallel systems to image processing provides the scaleability that is required to handle large real time applications. These systems can only be used efficiently with an architecture which is able to balance processing and memory with internal and external communication bandwidth. By efficiently harnessing the power of nodal heterogeneity and communication reconfigurability, systems with an ideal balance can be assembled and applied to achieve very high levels of efficiency. This paper outlines such a Heterogeneous Distributed Memory Parallel Processor (HDMPP) and discusses the software environment that is required to take advantage of it. Meiko's Application Center has considerable experience in applying these HDMPP systems to practical applications including: high definition television, radar, sonar, medical imaging, remote sensing, machine vision, data compression, image enhancement, and pattern recognition. These examples are used to illustrate how HDMPP systems have been successfully applied and how applications have been prototyped, developed, optimized and deployed on a single commercial off the shelf product.

To extract meaningful information from available data, researchers are often confronted with data in some complex superposed state. Therefore, the physical quantity of interest is not directly observable. In the physical sciences, a common form of information mixing is linear superposition. This includes fields as diverse as radio astronomy, Fourier transform spectroscopy, atmospheric physics, and medical diagnostics. One problem confronting researchers in these disciplines is restoring or deconvolving data. In addition to simple data restoration, noise can complicate the restoration process. This phenomenon can enter both prior to or during the collection of data. Noise presents a major obstacle to perfect restoration, information gain, and scientific understanding.

Parallel processing has been widely accepted as the approach to providing the necessary computational power to solve computer vision systems problems. Although several projects are underway to develop new architectures for computer vision, tools to effectively use those systems or commercially available multiprocessors are limited or non-existent. Unless we can develop efficient methods for mapping vision algorithms and developing programs on these architectures, the performance gains from parallel processing will be limited, and will be beyond the reach of a non-expert in parallel processing. This paper presents a design for a software development environment (SDE) for implementing vision systems applications on multiprocessors. The SDE design exploits characteristics of vision systems, and uses a classification scheme for vision algorithms to develop a parallelization and performance evaluation tool. These tools use databases that store knowledge of parallelization for different known computations on common architectures. The parallelization and performance evaluation tools use this knowledge to guide a user interactively parallelize algorithms. Some parts of SDE are currently operational and others still need to be developed.

Presented in this paper is a methodology that has been used in development of a model-based computer vision system. The methodology focuses on problem solving while using commercial off-the-shelf computer vision software.

A methodology and supporting software tools for obtaining insight and quantitative measures on the accuracy of ATC (automatic target cueing) algorithms are described. Algorithm accuracy is characterized by trade-off curves in which selected performance measures are plotted against each other over a range of algorithm operating points. Performance trade-off curves determined as a function of critical sensor, target and background parameters allow algorithm sensitivity issues to be investigated and provide a basis for evaluating algorithm accuracy effects in the context of specific missions. The evaluation process is founded on an ATC experiment design which defines a challenging but representative "cut" through the multidimensional scenario space associated with the mission under consideration. The experiment design is included in a written test plan which documents the mechanics of the evaluation process and specifies the amount and type of imagery that will be made available for algorithm development and evaluation. tolls for placing confidence bounds on hypothetical algorithm performance trade-off curves provide a basis for addressing the statistical data adequacy issue in relation to the experiment design.

The Strategic Computing Object-directed Reconnaissance Parallel-processing Image Understanding System (SCORPIUS) program was successfully completed in September 1990. Initiated in 1985, SCORPIUS was a research program that combined emerging technologies from DARPA's Image Understanding (IU) and Computer Architecture programs in a real world application, the automated exploitation of aerial imagery. Image exploitation is the process of extracting intelligence from image data. The quantity and quality of the products from various sensor systems are out-pacing the improvements in exploitation systems. SCORPIUS has demonstrated the concept that the model-based vision approach can address this imbalance.

The high volume of satellite derived oceanographic data, and the relatively high level of skill associated with the detection of important features in multi-sensor oceanographic datasets, has necessitated automating the analysis process. Since 1983 the Naval Oceanographic and Atmospheric Research Laboratory (NOARL) at NASA's Stennis Space Center has been involved in a effort which transitions research in automated interpretive techniques to operational use. The NOARL image understanding system's basic philosophy is unique in its strong emphasis on integration of different artificial intelligence techniques, conventional image processing techniques, statistical techniques, and low level vision techniques, making use of the strengths of each technique to optimally achieve the ultimate goal; an object based map showing icons which relate to detected features in the oceanographic imagery. The present paper describes the approach implemented, discusses lessons learned in past development efforts, and explains the rationale for the future evolutionary course of the system.

SCORPIUS is a model-based image understanding system developed on a suite of DARPAsupplied hardware. It detects and identifies objects in operational imagery over a range of imaging conditions. As a research project, SCORPIUS advances established technologies while exploring relatively young fields: image processing, image understanding, and parallel processing. During the five-year development of various system processing components, a substantial body of empirical knowledge was acquired.

A system was assembled to study several aspects of locating ship targets from infrared imagery. The system was either placed on shore sites or installed on an aircraft to collect data on the scene. The primary sensor was an infrared camera which produced images of the scene at standard RS-l70 rates. Requirements that included real time operation dictated the use of a parallel architecture for this task. As no suitable commercial systems were avail able, a custom array of bit slice microprocessors was assembled for the task. Through extensive field tests strengths and limitations of the design have been identified. These lessons are being applied to the development of next generation systems. A gimbal mounted infrared camera with digitization circuitry presents a new 256 by 256 pixel image to the parallel pipelined array of 17 bit slice microprocessors thirty times a second. To extend processor performance beyond the standard commercial microprocessors. two basic bit slice designs were employed. The bit slice machines were highly tuned for the assigned tasks and algorithms. Unfortunately this restricted the desired flexibility to readily examine alternate algorithms. The fundamental architecture concept performed well quickly reducing the large array of data to manageable set of information. Real time operator displays were driven to monitor the progress of each test run. Results of the system operation were stored on video and digi tal recorders permitting more detailed analysis after each test. Non real time data reduction provided many insights into the system operation and to algorithm improvements. Substantial operator interaction. and data interpretation was required greatly slowing the post test analysis phase. Overwhelmed with data, the analysts focused on locating a few data segments of interest. Significant work remains in improving the interfaces between the field data and the powerful laboratory computers. Automation of the data analysis is also needed to efficiently evaluate the great volume of field information. Continuing improvements in Artificial Intelligence, Expert Systems, Neural Networks, and other areas may help here.

In order to design vision systems which work, a sound engineering methodology must be utilized. In the systems engineering approach, a complex system is divided into simple subsystems and from the input/output characteristics of each subsystem, the input/output characteristics of the total system can be determined. Machine vision systems are complex, and they are composed of different algorithms applied in sequence. Determination of the performance of a total machine vision system is possible if the performance of each of the subpieces, i.e. the algorithms, is given. The problem, however, is that for most algorithms, there is no performance characterization which has been established and published in the research literature. Performance characterization has to do with establishing the correspondence of the random variations and imperfections which the algorithm produces on the output data caused by the random variations and imperfections of the input data. This paper illustrates how random perturbation models and propagation of random errors can be set up for a vision algorithm involving edge detection, edge linking, arc segmentation, and line fitting. The paper also discusses important dimensions that must be included in the performance characterization of any vision module performing a parametric estimation such as object pose, curve fit, or edge orientation estimation. Finally, we outline a general parametric model having three components: a relational model; a noise model; and a computational estimation model.