Multiple Model Particle Filtering For Multi-Target Tracking

Abstract : This paper addresses the problem of tracking multiple moving targets by recursively estimating the joint multitarget probability density (JMPD). Estimation of the JMPD is done in a Bayesian framework and provides a method for tracking multiple targets which allow nonlinear target motion and measurement to state coupling as well as non-Gaussian target-state densities. We utilize an implementation of the JMPD method based on particle filtering (PF) techniques. The details of this method have been presented elsewhere 1. One feature of real targets is that they are poorly described by a single kinematic model Target behavior may change dramatically i.e. targets may stop moving or begin rapid acceleration. To address this fact we evaluate the use of the adaptive target tracking strategy known as the interacting multiple model (IMM) algorithm. The IMM uses multiple models for target behavior and adaptively determines which model(s) are the most appropriate at each time step based on sensor measurements. We demonstrate the applicability of the IMM to a PF-based multitarget tracker in two settings. First we consider the traditional application of tracking targets that switch between kinematic modes. The target motion used is field data recorded during a military battle simulation and includes multiple modes of target behavior. Our investigation is distinguished from prior efforts in that it is concerned with multiple targets and real target motion data and utilizes a PF implementation. Second we present a nontraditional reinterpretation of the multiple model filter as multiple models on the state of the filter rather than on the state of the target. We find that this strategy is able to automatically detect model violations and compensate by altering the filter model which results in improved target tracking.