This research explores the classification capabilities of an automatic target recognizer operating in conjunction with a passive radar system. Our industrial partner in this work is the Silent Sentry development team at Lockheed Martin Mission Systems in Gaithersburg, MD. Because Silent Sentry is a commercially available passive radar, Lockheed Martin is able to inform us of both real-world difficulties and realistic operating characteristics of these sorts of tracking systems.
The following is a simple block diagram illustrating our classification approach.
We assume that the tracking radar supplies raw data in the form of measurements of reflected power from the unknown target of interest. In the passive radar scenario, the scattered signals might be TV or FM radio transmissions. The front-end of our classifier then processes the raw data to extract an estimate of the radar cross section (RCS) during each integrative interval. The actual pattern matching engine we use is Bayesian. As such, in the presence of equal prior probabilities, it seeks to evaluate the likelihood of the observed RCS vector for each potential class identity in the target library. Unfortunately, RCS is a function of the target's aspect with respect to both the illuminators and the receiver. (RCS also varies with the frequency and polarization of the incident wave). This unknown aspect is a nuisance parameter which must be dealt with in our Bayesian formulation. The aspect angle is fully specified by the target's position and orientation (e.g., roll, pitch, and yaw), since we already know the locations of the illuminators and the receiver. As such, in addition to reflected power measurements, our classifier also uses the position and velocity estimates provided by the radar (where velocity approximates orientation). Finally, it is worth mentioning that this unknown aspect angle typically will change from sample to sample as the target maneuvers.
In the course of computing the necessary likelihoods, we must determine expected RCS values for many different aspect angles at each sample instant. We accomplish this by interpolating between polarization matrices calculated using the Fast Illinois Solver Code (FISC) developed at the University of Illinois Center for Computational Electromagnetics . These polarization matrix look-up tables are generated off-line for each class in the target library. These tables form the "template library" indicated in the block diagram above.
Two examples of the template data produced by FISC are included at the left. Both panels are plots of bistatic RCS (in dBsm) derived from the HH scattering coefficient at 99.9 MHz. In both cases, elevation is set to zero. The x-axes correspond to incident azimuth; the y-axes correspond to scattered azimuth. The upper panel is the bistatic RCS of a Falcon-20 business jet; the lower panel is the bistatic RCS of a VFY-218 (an experimental aircraft). Pictures derived from CAD models of these targets are included at the bottom of this page.
In this work, it is one of our goals to implement a classifier which performs in real-time, or as close as possible. As such, we cannot afford to compute likelihoods over the entire aspect space spanned by the FISC tables in our library. Instead, we substantially speed up classification progress by making the following assumptions. Defining a target's unknown state as its three-dimensional position and angular orientation:
We have performed five-class Monte Carlo simulations using noisy data synthesized from an actual aircraft flight path. We also recently demonstrated the ability to successfully classify an unknown target in a four-class experiment using actual data supplied by Lockheed Martin (position/velocity estimates and reflected power). We are now investigating the performance improvement achievable by coupling the recognizer to the radar so that both influence each other as tracking and classification proceeds.
|Back to Top