This paper describes a new approach to automatic target recognizer (ATR) development utilizing artificial intelligent techniques. The ATR system exploits contextual information in its detection and classification processes to provide a high degree of robustness and adaptability. In the system, knowledge about domain objects and their contextual relationships is encoded in frames, separating it from low level image processing algorithms. This knowledge-based system demonstrates an improvement over the conventional statistical approach through the exploitation of diverse forms of knowledge in its decision-making process.
A human air traffic controller perceives possible aircraft collisions from a display of many aircraft. In the work described here, the computer understands the displayed aircraft conflict data by processing a global, semantic representation of the displayed data. Understanding implies that the computer can represent and interpret the displayed data in a manner suggestive of an experienced human controller. The representation is called the conflict structure. The paper describes the conflict structure and its use by an expert system that performs the enroute air traffic control task. An example is taken from a 'live' air traffic control training problem.
This paper describes a program of research designed to move a newly revitalized and promising approach to Computer Vision from the laboratory into the real world. In the first section, this "Expert Photo Interpreter" (EPI) approach is described and contrasted with two other approaches, which are termed "neural" and "physical." Next, the foundation and justification for this approach are presented, along with reasons to hope EPI may advance the state of the art in practical vision systems. Finally, a program of research to move the EPI approach from the laboratory into real systems is described.
The author and her team have developed a demonstration ocean surveillance information fusion expert system under an Internal Research and Development (IR&D) program at Science Applications, Inc. (SAI). Specifically, the expert system models the thought processes of an ocean surveillance watch analyst attempting to assess vessels' missions and destinations given their correlated tracks, history, and location/status of other vessels in the domain of interest. For the demonstration, rules were developed for determining the mission and destination of selected survey ships. The expert system was developed using the OPS5 production system and the Franz Lisp programming language on a VAX/VMS computer system. By the end of 1983, a simple demonstration was running. The development plans for 1984 include expanding the scope of the system to handle multiple vessel tracks and more mission types and enhancing the user interface of the system. The architecture and design of the expert system will be presented in this paper, implementation issues will be discussed, and progress and future plans will be overviewed.
Extending the recent successes demonstrated by artificial intelligence expert system technology to the broader domain of scene analysis necessitates a consequent broadening of the concepts and techniques used in current systems. Simple, single mechanisms must give way to multiple lines of reasoning and multiple levels of description. Meta-level reasoning, which is a recursive application of the basic expert system paradigm, is a promising approach to the problem of coping with the complexity inherent in highly variable, dynamic environments. This paper describes research directed towards incorporating meta-level reasoning into context-based scene analysis systems. A multi-layered expert system architecture is outlined that is aimed at providing high-level strategies and dynamic planning capability to the basic image understanding process.
This paper suggests an approach to modeling reactor control planning and decision-making which serves as a framework for the design of computer based decision aids. The suggested approach combines two distinct theoretical concepts. The first uses a probabilistic graph to characterize the operator's "internal model" of the system behavior; the second invokes the idea of implicit utility maximization as a generic computational heuristic operating on the graph. The model is used to derive some basic design principles of decision aids that might complement operator decision processes in planning control operations.
This paper describes a technique for systematically locating multiple faults in electrical and mechanical systems. The technique is based upon data flow programming, where the control of execution for a software test module is determined by the availability of data needed for the module to execute. The technique is applicable to automatic test units with small memory and processor throughput resources. Also, the technique is capable of supporting extensive CAD tools for the construction of the expert system.
The Lockheed Expert System (LES) has been designed to help knowledge engineers quickly solve problems in diagnosing, monitoring, planning, designing, checking, guiding, and interpreting. In its first application, LES was used to guide less-experienced maintenance personnel in the fault diagnosis of a large signal-switching network containing Built-In Test Equipment (BITE). LES used not only the knowledge of the expert diagnostician (captured in the familiar form of "IF-THEN" rules), but also knowledge about the structure, function, and causal relations of the device under study to perform rapid isolation of the module causing the failure. In addition to aiding the engineer in troubleshooting an electronic device, LES can also explain its reasoning and actions to the user, and can provide extensive database retrieval and graphics capabilities. In this paper we show how both the structure of the device and the troubleshooting rules of the expert are conveniently represented using LES's case grammar format. Also, an actual troubleshooting session between a user and LES is presented. By adding goals, rules, database information, and a few special procedures to the general LES framework, we were able to have a working system in a much shorter time (four man-months) than would have been possible starting afresh. The current status of the system is that it has been fielded and is under evaluation. LES is now being applied in other domains which include design checking and photo-interpretation.
This paper discusses some of the practical aspects of implementing expert systems in a real-time environment. There is a conflict between the needs of a process control system and the computational load imposed by intelligent decision-making software. The computation required to manage a real-time control problem is primarily concerned with routine calculations which must be executed in real time. On most current hardware, non-trivial AI software should not be forced to operate under real-time constraints. In order for the system to work efficiently, the two processes must be separated by a well-defined interface. Although the precise nature of the task separation will vary with the application, the definition of the interface will need to follow certain fundamental principles in order to provide functional separation. This interface was successfully implemented in the expert scheduling software currently running the automated chemical processing facility at Lockheed-Georgia. Potential applications of this concept in the areas of airborne avionics and robotics will be discussed.
A reasoning system for a practical case is presented. General knowledge about a universe of discourse is represented and processed in the system. The general knowledge implicitly represents concrete knowledge which due to its volume is impossible to represent explicitly. The formal and informal aspects of the used representation scheme are discussed. Further, the heuristic components of the reasoning system is explicated with respect to some of the identified processing requirements.
The Knowledge Acquisition System /Prospector expert system building tool developed by SRI, International, has been used to construct an expert system to identify aircraft on the basis of observables such as wing shape, engine number/location, fuselage shape, and tail assembly shape. Additional detailed features are allowed to influence the identification as "other favorable features." Constraints on the observations imposed by bad weather and distant observations have been included as contexts to the models. Models for Soviet and U.S. fighter aircraft have been included. Inclusion of other types of aircraft such as bombers, transports, and reconnaissance craft is straightforward. Two models permit exploration of the interaction of semantic and taxonomic networks with the models. A full set of text data for fluid communication with the user has been included. The use of demons as triggered output responses to enhance utility to the user has been explored. This paper presents discussion of the ease of building the expert system using this powerful tool and problems encountered in the construction process.
Phoneme hypothesization in continuous speech can be a difficult task, especially if speaker-independence should be achieved. In general, the interpretation of speech patterns involves the generation of hypotheses concerning possible phonemic transcription of syllable segments automatically extracted from a numerical representation of energy-time-frequency obtained by short-term spectral analysis of a spoken sentence. Each hypothesis is evaluated and a degree of worthiness is assigned to it in such a way that it can be further processed for hypothesizing words or syntactic or semantic structures of the sentence. Moreover, context dependencies among adjacent phonemes must be taken into account; for this purpose, segments of the order of the syllables are considered; in this way the context dependencies (at least in Italian) may only occurr within the speech unit selected. In this paper the method used for the recognition of stop sounds (/b/, /d/, /g/, /p/, /t/, /k/) is described; the system is organized as an expert system, in which various sources of knowledge cooperate. In particular, each expert contains a set of production rules, describing how the different phonemic hypotheses are related to phonetic or acoustic features.
A structure of preprocessing is described corresponding to a planner strata of the "perception-cognition" interaction within the machine intelligence for an autonomous mobile vehicle. A terraine is represented via a polygonal map. Algorithms are described which transform such a map into a database which can be used by PLANNER to solve motion planning problem.
McCarthy and Hayesl defined AI in terms of epistemology and heuristics. Research in natural language processing has added a third - a linguistic - dimension to their dichotomy. In this paper, I provide a discussion of general issues, related to natural language interfaces, as well as an overview of GROK2 (Grammatical Representation of Objective Knowledge), a natural language front end for the medical expert system MDX3.
Described is a general approach to the development of computer programs capable of designing image-forming optical systems without human intervention and of improving their performance with repeated attempts. The approach utilizes two ideas: 1) interpretation of technical design as a mapping in the configuration space of technical characteristics and 2) development of an "intelligent" routine that recognizes global optima. Examples of lens systems designed and used in the development of the general approach are presented, current status of the project is summarized, and plans for the future efforts are indicated.
The future implementation of a number of charge transfer device tracking and guidance sensors offers the opportunity to apply artificial intelligence to the problem of optimizing the tracking accuracy in a system whose optical performance is a function of time, field position, or source type. Intelligent sensors capable of optimizing their own performance can offer important reduction in the complexity and precision of the input optical system. The algorithm to be described uses multiple measurements of the optical line-spread function with known motion to define the parameters of a correction polynomial, which is applied to the raw centroid output. Several implementations of this algorithm are discussed, including hardware and software requirements, trade-offs between accuracy and speed, and ultimate performance. Results of a computer simulation of one implementation of the algorithm to Fine Guidance Sensor accuracy improvement of the Shuttle Infrared Telescope is described in detail. The effects of various optical defects on tracking performance with and without adaptive interpolator optimization will be shown. The use of the algorithm within the framework of a normal telescope is explained. The potential reduction in the wavefront quality of the imaging system is discussed.
There is rising interest in autonomous systems for future military environments. Systems that can take mission directives and then plan and conduct the desired mission will deeply involve aspects of machine intelligence for terrain navigation. Can we build such a system? The answer is "Yes, but definitely not today." Explanation of this answer requires reasons why today's technologies will not support autonomous vehicles and rationale for expecting these capabilities in the future. These reasons and rationale are presented by first defining functional requirements for an autonomous vehicle and its subsystems. Then the functional requirements are linked to technologies. The state of todays technologies and those expected in the future are outlined. Planned Army and DARPA programs in semi-autonomous and autonomous vehicles are described and shown to be the basis for future optimism for autonomous vehicles in the Army. Particular emphasis is given to terrain navigation concepts that will be generic for any type of autonomous ground vehicle.
This paper addresses the problem of route planning for ground vehicles. The problem is decomposed into two principal sub-problems: manipulation of a multi-dimensional knowledge base to result in a "composite map" consistent with the current mission goals, and a subsequent search procedure applied to this composite map to result in high performance routes. The relevance of expert systems and other techniques for route planning is discussed. A particularly efficient search procedure is applied to several example composite maps to demonstrate the power of the approach.
A system that performs automatic path planning for an autonomous land vehicle is described. It uses three levels of planning: a mission planner, a long range planner, and a local planner. The system relies both on a digital database and sensor-based information to plan a route, and it has been implemented as a software simulation for robotic vehicles.
Artificial intelligence planning systems attempting to achieve human-like performance typically bring to bear a wealth of real-world knowledge in order to select actions consistent with the system's goals and its assessment of the state of its environment. Unfortunately, as machine reasoning systems become larger and more general, they frequently become correspondingly slower and hence less effective at their intended task. Meanwhile, most human actors can deal competently with quite complex environments without compelling evidence that they plan by relying principally upon (or even understanding) formal reasoning and planning techniques such as resolution theorem proving, dynamic programming, and backward chaining. We suggest that humans can plan and replan so quickly because of two important principles: (a) their internal represention of the world is well suited to the planning problems they solve, and (b) their plans have much less depth than most powerful machine reasoning systems. A good substitute for deep planning may be a "broad but shallow" planning strategy that generates plans terminated in parameterized action sequences ("behaviors") which are chunked at a relatively high level of abstraction, combined with a context-dependent salience measure that differentially cues plan fragments or "behaviors" to propose themselves as candidates during time-critical planning operations.
This paper describes an autonomous airborne vehicle being developed at the Georgia Tech Engineering Experiment Station. The Autonomous Helicopter System (AHS) is a multi-mission system consisting of three distinct sections: vision, planning and control. Vision provides the local and global scene analysis which is symbolically represented and passed to planning as the initial route planning constraints. Planning generates a task dependent path for the vehicle to traverse which assures maximum mission system success as well as safety. Control validates the path and either executes the given route or feeds back to previous sections in order to resolve conflicts.
We present a simple heuristic-based navigation algorithm for autonomous vehicle con-trol. We assume sensor input in the form of a range scan of distances to closest obstacles at a variety of angles in the vehicle's field of view. The algorithm works within a 2-dimensional model of the world to solve the problem of obstacle avoidance : find a collision-free path from START to GOAL.
A concept of navigation is simulated based upon heuristic search. A mobile robot with a vision system navigates with an unknown or an unclear map. The range of vision is limited, thus, inflicting various judgments concerned with the comparison of alternatives of motion. The frequency of the decision-making procedure is limited by a definite time of computation. The system is simulated with a number of maps and the results of navigation are compared.
An algorithm for planning paths through a digital terrain map has been developed for guiding an autonomous land vehicle. A grid representation of the terrain allows finding optimal paths when considering terrain data such as elevation, mobility, cultural features, and military threats. A solution to the digitization bias of a grid representation is presented.
This research extends previous work on the modelling and analysis of time-varying imagery using distributed parameter systems (DPS) theory. An integrated or "weak" solution approach to the DPS model yields a computationally efficient algorithm for feature extraction and estimation of time-varying image motion which allows image function step discontinuities with respect to both spatial and temporal arguments. These algor-ithms represent a region-oriented as opposed to point-by-point approach to image motion analysis. Experimental results using real-world time-varying imagery confirm the validity of the approach. The particular effort reported herein concerns analysis of the spatio-temporal features extracted from the weak solution approach. Information related both quantitatively to motion estimation and qualitatively to motion and object(s) structure are shown to evolve from this approach. Thus, the blending of the "motion is basic" and "static is basic" approaches in dynamic imagery is enabled. To remove some of the "nearsightedness" of a parameter estimation approach to motion analysis, contextual motion analysis is employed for both extraction of dynamic scene regions and edge types and labelling of motion regions. The present status of each of these tasks is described.
Current automated approaches for recognizing spatially localized objects fall short of human recognition performance for two reasons. First, the classical statistical pattern recognizer uses only locally measured scene information in the recognition process. Contextual information can and should be extracted and exploited to enhance automated statistical approaches. Second, the human recognizer is able to learn from and adapt to new situations, while the automatic statistical pattern recognizer is confounded by any object that fails to match some subset of its training experience. Techniques from artificial intelligence can be employed to endow the machine with the ability to adapt to the scene at hand based on feedback of derived scene information. This paper describes an intelligent automated object recognition system that uses extracted scene context and feedback for improved recognition.
In this paper we present an interactive system which allows the Fourier analysis of the artichoke flower-head profile. The system consistsof a DEC pdp 11/34 computer with both a a track-following device and a Tektronix 4010/1 graphic and alpha numeric display on-line. Some experiments have been carried out taking into account some different parental types of artichoke flower-head samples. It is shown here that a narrow band of only eight harmonics is sufficient to classify different artichoke flower shapes.
The structural target analysis and recognition system (STARS) is a pyramid and syntactical based vision system that uniquely classifies targets, using their viewable internal structure. Being a totally structural approach, STARS uses a resolution sequence to develop a hierarchical pyramid organized segmentation and formal language to perform the recognition function. Global structure of the target is derived by the segment connectivity of the inter-resolution levels, while local structure is based on the local relationship of segments at a single level. The relationships of both the global and local structures form a resolution syntax tree (RST). Two targets are said to be structurally similar if they have similar RSTs. The matching process of the RSTs proceeds from the root to the leaves of the tree. The depth to which the match progresses before failure or completion determines the degree of patch in a resolution sense. RSTs from various views of a target are grouped together to form a formal language. The underlying grammar is transformed into a stochastic grammar so as to accommodate segmentation and environmental variations. Recognition metrics are a function of the resolution structure and posterior probability at each resolution level. Because of the inherent resolution sequence, STARS can accommodate both candidate and reference targets from various resolutions.
Graphical representation of cartocraphic data consists of symbols assigned to physical entities, interconnections to denote spatial and structural relationships among such symbols, and text to describe and identify these symbols and interconnections. An image understanding system to extract useful information from geographical maps is proposed. Pattern recognition techniques are applied to identify geometrical forms of various symbols. For each valid symbol identified, a set of files is created to store pertinent structural information extracted from the image. These files, in conjunction with a knowledge base of geographic data, are used to answer simple queries related to the spatial organization of the objects in the map.
Accurate and reliable detection of unique objects is an important component of an image understanding system. The objects are considered to have a unique pattern and should be recognized based upon their own characteristics. An object detection approach applicable to the high resolution aerial images is developed in the paper. The methodologies incorporate features such as minimal sensitivity to the backgrounds on which objects may appear, ability to detect objects appearing in arbitrary orientations, and use of a unified set of operators to analyze various levels of details. Specifically the gray level cooccurrence matrix is used to provide the early vision operators. Results of several experiments show the promise for the proposed methods.
Shannon's theory of information is used to optimize the optical design of sensor-array imaging systems which use neighborhood image-plane signal processing for enhancing edges and compressing dynamic range during image formation. The resultant edge-enhancement, or band-pass-filter, response is found to be very similar to that of human vision. Comparisons of traits in human vision with results from information theory suggest that: (1) Image-plane processing, like preprocessing in human vision, can improve visual information acquisition for pattern recognition when resolving power, sensitivity, and dynamic range are constrained. Improvements include reduced sensitivity to changes in light levels, reduced signal dynamic range, reduced data transmission and processing, and reduced aliasing and photosensor noise degradation. (2) Information content can be an appropriate figure of merit for optimizing the optical design of imaging systems when visual information is acquired for pattern recognition. The design trade-offs involve spatial response, sensitivity, and sampling interval.
A new consulting system using a natural language and a graphical interface is under construction to assist a naive user in decomposing and constructing a mechanical object with cylindrical bodies. Many trouble shooting systems have been developed so far, but most of them do not tell us the way for decomposing the object to find out trouble points. This system is built to assist naive user in decomposing a mechanical object and in constructing it after repairation. It is difficult for a computer to give him a series of operations necessary for exposing a trouble point by using just simple command suquences, then an integrated instruction facility using a natural language and a visual interface must be offered to users for specifying what portion of the object should be decomposed or constructed at the next stage, and for verifying whether what the user have done to the object is correct or incorrect. The present art of computer vision cannot verify if an act taken by the user is correct or not at each step, because mechanical objects sometimes have involved structures. This system leaves this verification process to the user by showing him two perspective views of the objects, and an explanation on the operation which causes these two views before and after decomposition or construction.