Assignment-based particle labeling for PHD particle filter

Abstract

The probability hypothesis density (PHD) filter is an estimator that approximates, on a given scenario, the multitarget distribution through its first-order multitarget moment. This paper presents two particles labeling algorithms for the PHD particle filter, through which the information on individual targets identity (otherwise hidden within the first-order multitarget moment) is revealed and propagated over time. By maintaining all particles labeled at any time, the individual target distribution estimates are obtained under the form of labeled particle clouds, within the estimated PHD. The partitioning of the PHD into distinct clouds, through labeling, provides over time information on confirmed tracks identity, tracks undergoing initiation or deletion at a given time frame, and clutter regions, otherwise not available in a regular PHD (or track-labeled PHD). Both algorithms imply particles tagging since their inception, in the measurements sampling step, and their re-tagging once they are merged into particle clouds of already confirmed tracks, or are merged for the purpose of initializing new tracks. Particles of a confirmed track cloud preserve their labels over time frames. Two data associations are involved in labels management; one assignment merges measurement clouds into particle clouds of already confirmed tracks, while the following 2D-assignment associates particle clouds corresponding to non-confirmed tracks over two frames, for track initiation. The algorithms are presented on a scenario containing two targets with close and crossing trajectories, with the particle labeled PHD filter tracking under measurement origin uncertainty due to observations variance and clutter.