Principles of Modern Radar - Volume 2 1891121537

678 CHAPTER 15 Multitarget, Multisensor TrackingFeature-assisted tracking (summarized in Section 15.2.1.1) has the potential to improvemeasurement-to-track association when the kinematic costs across several elements of thecost matrix are similar; it augments the kinematically-based assignment costs with featurebasedcosts. Regardless of whether they incorporate features, sophisticated techniques forsolving the cost matrices to find optimal assignments, such as multiple hypothesis trackers(MHTs), are sometimes required to maintain sufficiently pure tracks in the presenceof dense target scenes. Although these filters incur significant computational costs, theyprovide a viable option if track purity is required in the context of a dense target scene.The MHT concept is summarized in Section 15.2.1.2. An alternative approach is summarizedin Section 15.2.1.3; rather than try to resolve measurement-to-track data associationambiguity, cluster tracking algorithms embrace it and track a set of closely-spaced objectsas a group (i.e., a cluster), hence reducing the computational requirements and necessarycommunications bandwidth.15.2.1.1 Feature-assisted TrackingWhen several elements of the kinematically-based measurement-to-track cost matrix aresimilar, the ‘correct’ association becomes ambiguous. One technique for resolving thisambiguity is to augment the kinematically-based cost function from (15.7) with a featurebasedterm. If the kinematics and feature are assumed to be independent (a commonassumption to simplify the math, even though it is not really true), then the total cost issimply the product of the kinematic cost and feature cost.Literature proposing techniques for feature-assisted tracking is abundant [22, 25, 41],some of which focuses on features that are difficult to routinely collect (due to limitedradar resources) and much of which focuses on ground moving target indication (GMTI)environments with a limited number of targets. With regards to GMTI, feature-assistedtracking is a useful tool for resolving ambiguities associated with target crossings; comparisonof time-invariant features before and after the crossings (e.g., of vehicles at anintersection) can mitigate track swaps. Rather than focusing specifically on GMTI applicationswith relatively small numbers of targets, or on feature-assisted tracking usingfeatures that can rarely be collected en masse, this section focuses on recent researchapplying signal-to-noise ratio (SNR)-assisted tracking to dense target scenes.Since feature-assisted tracking is often applied in conjunction with an MHT to addressscenarios with ambiguous assignment, this discussion uses the full hypothesis cost fromMHT literature, given byc i, j =−ln [ P( k.l |Z k ) ] (15.19)where P( k.l |Z k ) is the probability of hypothesis l (a particular pairing of measurementsand tracks) at time k. Prior work [4] derives this to be given byP( k.l |Z k ) = γm(k)∏j=1{ [fi, j z j (k) ]} τ j∏ [PDt (k) δ i(1 − P Di (k)) 1−δ ] iP( k−1.s |Z k−1 )i(15.20)where k.l is the joint association of all measurement/track pairings in hypothesis l att k , Z k is the cumulative set of measurements at t k , f i, j is the probability density function(pdf) of the residual for track i and measurement j, P Di is the probability of detection fortrack i, τ j is the indicator function for the assignment of measurement j to any track inhypothesis l, and δ i is the indicator function for the assignment of any measurement to