Performance Analysis of Sequential Monte Carlo MCMC and PHD Filters on Multi-target Tracking in Video

Abstract

The Bayesian approach to target tracking has proven to be successful in the tracking of multiple targets in various application contexts. This paper applies sequential Monte Carlo (SMC) filtering techniques such as the Markov Chain Monte Carlo particle filter (MCMC PF) and the SMC probability hypothesis density (PHD) filter as suboptimal Bayesian solutions to multi-target tracking (MTT) in video. The MCMC PF by virtue of its information-centric property, can automatically explore the posterior distribution at each sampling step making it possible to track multiple targets. In doing so, it propagates the full multi-target posterior. The SMC PHD filter however propagates only the first order moment of the multi-target posterior density thereby making it computationally less intensive. A comparison of both filters was carried out in tracking multiple human targets in a video scene demonstrating superior performance by the SMC PHD filter in a realistic scenario. The SMC PHD filter was seen to have higher performance than the MCMC PF in terms of the number of particles, the processing speed, and the tracking performance for multiple targets.