Principles of Modern Radar - Volume 2 1891121537

694 CHAPTER 15 Multitarget, Multisensor Tracking15.3.4 Sensor Fusion Challenges Common to Both ParadigmsThe aforementioned challenges are unique to measurement-level or track-level fusion, butseveral additional challenges are common to both fusion architectures. As was previouslydiscussed, sensor fusion is fundamentally a function of the contributing reports’ uncertainties.Measurement-level fusion is heavily influenced by reports with small measurementcovariances, and track-level fusion is similarly influenced by reports with small track covariances.One significant challenge is that sensors and trackers are often poor judges oftheir uncertainties, upon which fusion depends. They typically account for known biases(via calibration), but these address only the known-unknowns. Calibration cannot addressthe host of unknown factors that increase error, or the unknown-unknowns. Hence, radarsfrequently under-report errors. Techniques for improved bias estimation are widely availablein the literature, but application of ‘fudge factors’ that inflate reported covariancesare often used in practice, even though they are sub-optimal.Depending on the application, the sensors whose results are being fused may bedeployed over a large geographical space. If this is the case, then gaps in sensor coveragepose another challenge for both fusion paradigms. Situational awareness (a common goalfor multi-sensor fusion systems) mandates that the system track set (i.e., the result offusion in either scheme) include one track per reported object. If the objects all exit thecoverage area for one set of sensors before entering the coverage region of another, thisimplies that both sets of tracks must be fused to create the desired single system track perobject. This, in turn, requires the states and covariances to be propagated to a commonreference time. Depending on the sensor laydown, this may translate into tens or hundredsof seconds of propagation. The challenges compound. Fusion of perfectly unbiased, fullysettledtracks with perfectly consistent covariances may not be terribly degraded by longpropagation times, but long propagation times will severely degrade fusion results if thetracks are biased, are short-lived, or have inconsistent covariances. Propagation schemesthat identify the optimal time frame in which to conduct association may mitigate thischallenge [16].The deployment of disparate radars further compounds these problems. Radars withdifferent sensitivities, range resolution, viewing aspects, and coverage are unlikely to reporton identical objects from the threat complexes. In fact, the subsets of objects that each reportmay not overlap significantly. Notionally, one sensor with poor sensitivity and resolutionmay report a handful of tracks from a given scene, while another with excellent sensitivityand resolution may report hundreds of tracks on the same set of objects. Association andfusion become extremely challenging in this context, sometimes referred to as ‘subsetconditions’, since the subsets of objects being tracked by the different radars may notsubstantially overlap. Attempting bias estimation and removal in this context (if divorcedfrom physics-based rationale) can be treacherous, as numerous, non-unique biases maycause the small set of tracks to align well with subsets of the larger one, at least for shortperiods of time.Whether the contributing radars report discrimination feature observations (as wouldbe likely in a measurement-level fusion scheme) or track-based discrimination answers(in the track-fusion paradigm), fusing discrimination answers from multiple sensors isextremely problematic. Just as disparate sensors at differing deployment locales will reportdifferent numbers of tracks on the same complex, they may also collect different features(with different levels of accuracy) with different intended uses. While work has begun inthis area, it is ripe for further research.