Copyright 1990 Horizon House Publications Inc. Journal of Electronic Defense July, 1990 SECTION: Vol. 13 ; No. 7 ; Pg. 38; ISSN: 0192-429X LENGTH: 3342 words HEADLINE: Target recognition for autonomous smart munitions. BYLINE: Cohen, Marvin H. BODY: The phrase "smart munitions" refers to weapons that, most generally, are meant to acquire, track, target and kill potential threats. The sensing and signal processing required to actuate these functions are contained in the seeker portion of the smart munition. One subclass of smart munitions consists of those weapons that, once targeted and fired on a specific platform, will track and kill the platform either autonomously or with the aid of outside intervention. These smart munitions, such as the terminally guided Hellfire air-to-ground missile and the Advanced Medium Range Air-to-Air Missile, require no autonomous target recognition capability, because the target is selected and designated by an independent source. The class of smart munitions requiring inclusion of automatic noncooperative target recognition (NCTR) is the autonomous smart munition. The characteristics that distinguish this class from other types of seekers include (1) they are usually fired into an area where potential targets are thought to reside rather than at a specific target and (2) they are truly fire-and-forget, autonomous weapons. In general, these munitions are being considered for the Strategic Defense Initiative (SDI) as well as conventional air-to-ground and ground-to-ground applications. Typically, many such munitions are delivered to the general area of potential targets within a single bus," which can be an aircraft, artillery shell or guided missile. Once in the proper area, the munitions are released from the bus and are left to detect and destroy high-value targets on their own. Autonomous smart munitions have great tactical value because they allow the targeting of valuable assets well behind the forward line of troops while, because of their fire-andforget nature, preserving the security and flexibility of the launching platform. Much of the smarts in autonomous smart munitions resides in the munition's capability to detect targets autonomously, discriminate targets from clutter and prioritize targets according to type, status and identity. That is, central to the function of smart munitions is their ability to perform the functions of NCTR automatically and autonomously. Concomitantly, autonomous smart munitions must be low priced, small in size, immune to mutual interference and relatively immune to countermeasures. This article will focus on the sensors and signal processing required for implementation of effective NCTR in autonomous smart munitions, keeping in mind the constraints mentioned above. Figure 1 provides an artist's conception of the operation of a generic Sense and Destroy Armor (SADARM) autonomous smart munition. The snapshot represents a moment in time after the sensor fused munitions (SFMs) have been released by the bus and they are surveilling the local area. The SFMs autonomously detect potential targets, sort and prioritize them and finally destroy the appropriate targets by firing a self-forging fragment warhead. SENSORS The sensors of principle interest for autonomous smart munitions are radiometry, radar and infrared (IR). Radar provides all-weather, relatively long range capability. IR and radiometry provide passive day/night capability. Many systems applications seem to require dual-mode seekers to reduce false alarm rates and to provide continued effectiveness in the face of countermeasures. In particular, few natural objects tend to cause false alarms in both the radar and IR sensor domains; that is, since the false alarms in the two domains tend to be uncorrelated, radar and IR can be expected to greatly reduce false alarms when the data from the two sensors are fused. Furthermore, countermeasures in either the IR or radar domain tend not to affect the performance of the other. Thus, we will focus the discussions that follow on a dual-mode, radar, IR seeker. In radar, beamwidth (for a given antenna aperture) and component size are both proportional to wavelength. Thus, seeker designs tend toward the high-frequency, millimeter-wave (MMW) band. In particular, the combination of small wavelength and atmospheric absorption characteristics tend to push such designs into the 35-, 95-, or 140-GHz region of the electromagnetic spectrum. Depending on the specific mission, the radar sensor must have the capability to generate high range resolution (HRR) and/or broad Doppler coverage and fine Doppler resolution. HRR requires transmission of a wideband waveform and allows the determination of fine physical features of the potential target. Broad Doppler coverage allows the detection and mapping of fast- moving parts on the potential target and requires a high pulse repetition frequency (PRF) waveform. Fine Doppler resolution provides the ability to analyze specific moving components and requires relatively long time-on-target. Thus, depending on the mission, an autonomous smart munition's radar may well have to be a multimode, coherent, MMW radar that is relatively inexpensive. The design of such a system can represent a significant challenge. The IR sensor usually operates in wavelength bands. Historically, the approach has evolved from a single-band (such as 8 to 12 microns), single-pixel detection process to a dual-band (e.g., 8-12 and 1-3 microns), single-pixel detection process, through single-and dual-band, imaging detection and recognition processes. If the IR's function is simply as an aid in detection to reduce false alarm rates, a simple, single-pixel detector may be implemented. If, on the other hand, the IR sensor is required to play a major role in the more sophisticated NCTR functions of the munition, then an imaging IR seeker must be implemented. A critical factor in the development of an effective dual-mode seeker is the ability to align the received signatures from the two domains so that there is a clear positional correspondence between them. This is known as the scene registration problem, and it plays a central role in sensor fusion research. NCTR FUNCTIONS The functions of recognition may be conceptualized as stages in a recognition processor, as depicted in Figure 2. The techniques implemented for each of these stages differ significantly depending on whether the phenomenology one is trying to exploit is contained in the structure of moving or stationary parts of the potential target. If one can be sure that the potential target is moving (or some part of it will be moving, such as a scanning radar dish), then there may be exploitable features in the Doppler radar signature of the target. Otherwise, such as in the case of a stationary tank with its ignition turned off, one must rely on HRR profiles and images to provide the required features. In the discussion that follows, both these cases are discussed. As depicted in Figure2, the most fundamental level of recognition is detection, where potential targets are distinguished from thermal noise. Detection of moving targets is accomplished on the basis of moving target indication (MTI) techniques through the radar system and through the superposition and cancellation of sequential images through the IR system. In both cases, good detection performance is a function of the signal-to-noise ratio that one can achieve in the target signatures. The radar system must either be coherent with good motion compensation capability or it must employ some innovative technique such as clutter-referenced MTI. The IR sensor requires good motion compensation and at least a rudimentary image processing capability to perform image-to-image cancellation. In both cases, adequate detection capability is a function of the energy one can collect from the target which, in turn, is dependent on the system's hardware and operational parameters. The detection of stationary targets is accomplished in much the same way as above. However, the sequential IR images would be added rather than differenced, and the detection process will do little if anything for separating target returns from background clutter returns. Discrimination, the process of distinguishing potential targets from surrounding clutter, is a particularly important element of any recognition system. The discrimination processor (or process) must reduce the number of potential targets presented to the higher levels of the recognition system to minimize throughput requirements and to ensure good recognition performance. Form any high-speed targets in many scenarios, classical MTI techniques and sequential image differencing significantly reduce the difficulty of discriminating moving targets from stationary clutter. However, the autonomous smart munition scenario often requires the discrimination of slow-moving and stationary targets from background clutter in a severe look- down geometry. The classical techniques for discriminating slow movers and stationary targets from clutter generally rely on target amplitude. In radar, this refers to the radar cross section being large enough to distinguish the target from the surrounding clutter, and in IR this refers to the existence of target "hot spots" (such as engines) and/or emissivity contrast between the target body and the clutter. In both cases, resolution cell sizes are matched to target sizes to maximize in-cell target-to-clutter ratios. Thereafter, amplitude-based, area constant false alarm rate processors are used to meet the challenge of target/clutter discrimination when the clutter is relatively homogeneous. In nonhomogeneous or extremely high homogeneous clutter, more innovative approaches are required. Polarimetric properties, phase response, length estimators and complexity estimators have been applied to radar signatures for this purpose. Shape and dimension estimators as well as two-band operation have been applied in the IR domain for the same purpose. In general, the ability to premeasure and store ground maps in the autonomous smart munition signal processor - for example, to determine the location of roadways - can be another important asset to the accomplishment of target/clutter discrimination in the autonomous smart munitions system. This approach, however, can be quite difficult to implement in tactical scenarios with the IR sensor because of the extreme variations in scene signatures as functions of weather, time of day and time of year. The next level of recognition is often termed preclassification, or alien separation. The higher levels of recognition are almost always implemented via algorithms that are "trained" to recognize particular targets (or classes of targets). The library of training signatures or training features associated with each target of interest is, by necessity, severely limited as compared with all the possible signatures that the autonomous smart munitions system may encounter in the field. For example, one might like to prescreen targets such as automobiles, buses and farm tractors before presenting the incoming data to an algorithm trained to classify among tracked military vehicles and low-flying helicopters. The function of the preclassifier in any such system is precisely for recognizing and excluding from further consideration those targets that may a ppear in a scene but are not of sufficient interest to have been the subject of an intense analysis and algorithm training exercise. The term "alien separation" is highly suggestive of the function of this level of recognition; that is, prescreen or "separate-out" those targets that are not of the types of interest for further recognition and thus have not been used to train the classification or identification algorithms. Preclassification is often accomplished via the implementation of clustering techniques, where the features used are ones for which the targets of principle interest have similar values. The algorithm would then determine a cluster defined by these features, and would exclude from further consideration all potential targets that do not fall sufficiently close to this cluster. Thus, sensor detections whose characteristics are not sufficiently similar to those of all targets of interest are excluded. Classification represents a rather sophisticated level of the recognition process where potential targets are characterized according to some class groupings, such as wheeled versus tracked vehicles. The class groupings must have two important qualities. First, the groupings must be useful. For example, the class groupings of wheeled versus tracked vehicles can be quite useful in determining the threat potential of the incoming target as well as the correct choice of munitions to defeat it. Second, the groupings must entail physical attributes (characteristics) that distinguish members of different groups as well as attributes that are common to members of each group, and the sensor must be able to reliably measure these attribute differences and commonalities. The first quality relates to the usefulness of the classification process. The second quality relates to its viability. Typical classification functions might include fixed-wing versus helicopter, tracked versus wheeled vehicles or surface-to-air missile site versus anything else. In each case, it is the nature of the potential threat that is being determined rather than its specific identity or even whether it is friend or foe. Classification algorithms are generally developed by instituting "training" or "learning" process. In such a process, measured or simulated data of the targets of interest are analyzed and then used to provide thresholds or parameter values for the classification algorithms. Alternatively, these data may be submitted directly to the algorithms so that the algorithms adapt these optimal thresholds and parameter values themselves. In ground-to-ground tactical autonomous smart munitions systems, where the principal targets of interest are usually military tracked and wheeled vehicles and low-flying helicopters, it is likely that one might want to institute a classification algorithm to distinguish among these three classes of targets. Another important application of classification in the autonomous smart munitions arena is the determination of whether a given platform has already been made inoperable. Distinguishing tanks in the field that have been hit and destroyed by munitions from those not yet assaulted is an important classification function for the autonomous smart munition's recognition process. The features used to perform classification are usually finely detailed to characterize the targets. Thus, depending on the scenario, HRR profiles and/or high-resolution Doppler signatures might be used in the radar domain, while some form of imaging and image understanding techniques would almost certainly be applied in the IR domain. Identification is typically thought of as the most sophisticated and most difficult of the recognition functions. Identification algorithms are developed using methodology similar to that employed in the development of classification algorithms; that is, the algorithms are trained on a representative subset of the targets of interest. However, identification refers to the ability to specify a platform as, for example, an M-60, T-72, etc., rather than simply classifying an incoming target as a wheeled or tracked vehicle. More generally, the identification process may refer to any one of a number of various functions such as identification-friend-or-foe (IFF) or identification of the specific platform (T-72, Serial Number 223, etc.). The identification function employed, as with classification, depends on both utility and viability within the context of the available sensors and the operational scenario. Thus, although successful IFF will provide high utility for almost any scenario, its viability is much higher if one is in a scenario where it is known that the foe is not using the same equipment we are, than it would be if the current foe were a onetime friend, using domestically manufactured equipment against us. MUTUAL INTERFERENCE AND COUNTERMEASURES The philosophy behind the use of certain autonomous smart munitions includes the fact that because they are relatively inexpensive, one can afford to deploy many of these seekers for each target. In addition, it is highly likely that as these systems are deployed, active countermeasures will be fielded as well. As a result, it is expected that these systems will have to operate in a relatively dense electromagnetic environment. If care is not taken, mutual interference among the radars of each of the munitions can have a disastrous effect on performance. To overcome this potential problem, a band of radar frequencies can be made available to reduce the probability that two seekers using the same frequency will mutually jam each other. Also, the use of relatively orthogonal pulse compression codes ratherthan simple pulse or FMCW waveforms can go a long way toward reducing the potentially devastating problem of RF mutual interference. Countermeasures can include both passive and active techniques. Passive techniques such as camouflage, the addition of energy absorbing material and superficial structural modifications can pose substantial problems in the discrimination, classification and identification of stationary ground vehicles. Algorithms developed on the basis of the Doppler spectra of the intrinsic dynamics of certain moving targets may be much more difficult to defeat through the use of such passive countermeasures techniques. Active techniques include jamming of the radar waveform, the laying down of IR obscurants and the deployment of flares. The autonomous smart munition's radar can be hardened against jamming via the use of the frequency diversity and waveform encoding techniques mentioned above. When these techniques fail to provide the requisite protection, antiradiation missiles may be required to silence the jammer. While obscurants and flares can greatly reduce the effectiveness of passive IR sensors, they do little to inhibit the radar sensor in a dual-mode system. In fact, their use may provide valuable cues that could allow more effective use of the MMW radar sensor. SUMMARY Autonomous smart munitions show promise as an effective force multiplier in conventional tactical scenarios and as a longrange threat neutralizer in strategic scenarios. Much of the capability of such munitions depends on the ability to detect potential targets autonomously, discriminate them from surrounding clutter, separate them from benign detections, classify them according to their value and vulnerability and identify them as foes. The ability to execute these functions is greatly enhanced by the use of dual-mode seekers in conjunction with sophisticated fusion and NCTR signal processing and decision-making algorithms. Principal challenges in the field include effective scene registration between the sensors' fields of view; development of effective, robust NCTR algorithms in the presence of countermeasures; and the cost-effective implementation and packaging of such systems for field deployment. Acknowledgement: Portions of this article have been abstracted from the author's chapter, "Target Recognition in Airborne Early Warning Systems," to appear in Dr. Maurice Long's new text, Airborne Early Warning System Concepts.