Comparison of Two Particle Filter Methods For Image Plane Tracking of Multiple Objects

Abstract

This paper examines the capabilities of the Particle Filter (PF) for the tracking of multiple objects moving on a focal plane array image plane. The study considers the tracking of four objects for which the dynamics are near constant velocity with random perturbations in acceleration. The filtering structure uses four separate PFs rather than a concatenated filter with four sets of target states. There is a data association problem in which each measurement must be assigned to the proper PF. A new association logic based on weighted nearest neighbor strategy is employed to solve this problem. Two new methods are considered for calculating the PF importance weights. The first is based on a weighted distance between the PF position estimate and the perceived location of maximum measurement intensity. The second method employs a likelihood ratio to compute the weights. The first method performed well at a signal-to-noise ratio (SNR) of 15 dB or higher but was unsatisfactory below this. The second method was found to perform well at very low SNR approaching 0 dB. The paper discusses the multiobject tracking scenario, describes the association logic and PF implementations, discusses the analysis details, and shows the performance obtained from the PF.