MAP Estimation in Particle Filter Tracking

摘要

Posterior densities in nonlinear tracking problems can successfully be constructed using particle filtering. The mean of the density is a popular point estimate. However, especially in multi-modal densities it does not always represent a reasonable estimate. In multi-target tracking the mean can produce a large bias when there is uncertainty about the labelling of the tracks, also referred to as the mixed labelling problem. The particle based Maximum A Posteriori (MAP) point estimator that has been recently developed is applied to this problem. It is shown by means of simulation that it provides a large improvement over the mean estimate.