Evolutionary MCMC Particle Filtering for Target Cluster Tracking
Avishy Y. Carmi, Simon Godsill, François Septier · 2009
A new filtering algorithm is presented for tracking multiple clusters of coordinated targets. Based on a Markov chain Monte Carlo sampling mechanization, the new algorithm maintains a discrete approximation of the filtering density of the clusters' state. The filter's tracking efficiency is enhanced by incorporating two stages into the basic Metropolis-Hastings sampling scheme: 1) Interaction. Improved moves are generated by exchanging genetic material between samples from different realizations of the same chain, and 2) Optimization. Optimized proposals in terms of likelihood are obtained using a Bayesian extension of the EM algorithm. In addition, a method is devised based on the Akaike information criterion (AIC) for eliminating fictitious clusters that may appear when tracking in a highly cluttered environment. The algorithm's performance is assessed and demonstrated in a tracking scenario consisting of several hundreds targets which form up to six distinct clusters in a highly cluttered environment.