Extensions to the Probabilistic Multi-Hypothesis Tracker for Improved Data Association
Samuel J. Davey · 2003
Multitarget tracking is a state space estimation problem where false measurements, missed detections, and uncertainty in the source of measurements provide the challenge. The Probabilistic Multi-Hypothesis Tracker (PMHT) is an algorithm which solves the Multitarget tracking problem through application of the Expectation Maximisation algorithm. This algorithm has a number of advantages over traditional techniques, but has not undergone the same degree of development as more established algorithms. This thesis presents extensions to the PMHT which both generalise its fundamental problem formulation, and address practical issues arising in the use of real sensors. The PMHT is extended to incorporate augmented measurements, which consist of the normal state observations, and classification measurements not considered under the standard PMHT. These classification measurements are interpreted as observations of the assignments, and a PMHT algorithm is derived. The classification measurements improve data association, and simulations are used to demonstrate the effect this has on state estimation accuracy. The probabilistic assignment model, central to the PMHT, is generalised to allow for an assignment prior distribution which varies smoothly with time. The prior is modelled as a random process following a first order Markov chain. Simulations are used to demonstrate the performance of the algorithm under a time evolving assignment prior. The PMHT assumption of a constant and known number of targets is relaxed by developing automatic track initiation schemes which are used to reject superfluous candidate models. Several approaches, similar to those used in other tracking algorithms, are considered, and the best of these is found by simulations involving various clutter conditions. The above extensions are applied to the problem of Over the Horizon Radar (OTHR) tracking and a prototype OTHR tracker is developed. A number of OTHR specific problems are also addressed, and the performance of the PMHT extensions is measured on data recorded from an operational OTHR. The performance of the PMHT prototype is compared with the existing tracking algorithm, which is based on the Probabilistic Data Association Filter.