《Second generaztion PHD or CPHD filters and multitarget calculus》.pdf

“Second-generation” PHD/CPHD filters and multitarget calculus Ronald Mahler Lockheed Martin MS2, Eagan MN ABSTRACT The “classical” PHD and CPHD filters presume the standard “small-target” detection model. This year, in a series of theoretical studies, I have derived new “second-generation” CPHD/PHD filters for various sensing conditions that cannot be described by the standard model. These are: (a) multisensor, (b) clutter estimation, (c) tracking in unknown clutter, (d) extended targets, (e) unresolved targets, and (f) superpositional sensors. A common factor underlies all of these derivations: the FISST multitarget calculus. It is possible, given that one already knows the correct “answer,” to reverse engineer the classical PHD/CPHD filters and to extemporize some “elementary” means of deriving them. But only the multitarget calculus is guaranteed to result in theoretically rigorous formulas for new problems–i.e., those for which the answer is not known beforehand. I also announce an important new result: the multitarget state estimators used with the CPHD/PHD filters are Bayes-optimal. 1 Introduction In 2000, I introduced a multitarget detection and tracking approach called the probability hypothesis density (PHD) fi lter.8 In place of the full multitarget probability distribution fk |k (X |Z(k)), this filter propagates a multitarget statistical fi rst moment–the PHD Dk |k (x |Z(k)). The PHD is (in general) not a probability density, since its integral Dk |k (x |Z (k))dx is the expected total number Nk |k of targets in the scene. In 2006 I introduced a generalization called the cardinalized PHD (CPH