Unscented Auxiliary Particle Filter Implementation of the Cardinalized Probability Hypothesis Density Filter

TL;DR

This paper introduces an unscented auxiliary particle implementation of the CPHD filter, improving multi-target tracking accuracy in nonlinear, cluttered environments compared to existing SMC-based methods.

Contribution

It proposes a novel unscented transform-based auxiliary particle filter for the CPHD, enhancing accuracy and efficiency in nonlinear multi-target tracking.

Findings

Significantly outperforms SMC-CPHD in accuracy

Better handles high clutter scenarios

Improves run time and target appearance sensitivity

Abstract

The probability hypothesis density (PHD) filter alleviates the computational expense of the optimal Bayesian multi-target filtering by approximating the intensity function of the random finite set (RFS) of targets in time. However, as a powerful decluttering algorithm, it suffers from lack of the precise estimation of the expected number of targets. The cardinalized PHD (CPHD) recursion, as a generalization of the PHD recursion, is to remedy this flaw, which jointly propagates the intensity function and the posterior cardinality distribution. While there are a few new approaches to enhance the Sequential Monte Carlo (SMC) implementation of the PHD filter, current SMC implementation for the CPHD filter suffers from poor performance in terms of accuracy of estimate. In this paper, based on the unscented transform (UT), we propose an auxiliary implementation of the CPHD filter for highly nonlinear systems. To that end, we approximate the elementary symmetric functions both with the predicted and with the update estimate of the linear functional. We subsequently demonstrate via numerical simulations that our algorithms significantly out performs both the SMC-CPHD filter and the auxiliary particle implementation of the PHD filter in difficult situations with high clutter. We also compare our proposed algorithm with its counterparts in terms of other metrics, such as run times and sensitivity to new target appearance.