Multi-Target Tracking Using A Randomized Hypothesis Generation Technique

TL;DR

This paper introduces R-FISST, a randomized Bayesian filtering approach for multi-object tracking that employs MCMC sampling to efficiently handle large hypothesis spaces, demonstrated on space situational awareness data.

Contribution

The paper develops a randomized FISST method that makes multi-object tracking computationally feasible using MCMC sampling, unifying FISST and MHT approaches.

Findings

R-FISST effectively tracks 50 objects in SSA scenarios.

R-FISST outperforms HOMHT in accuracy and computational efficiency.

The approach is scalable to complex multi-target tracking problems.

Abstract

In this paper, we present a randomized version of the finite set statistics (FISST) Bayesian recursions for multi-object tracking problems. We propose a hypothesis level derivation of the FISST equations that shows that the multi-object tracking problem may be considered as a finite state space Bayesian filtering problem, albeit with a growing state space. We further show that the FISST and Multi-Hypothesis Tracking (MHT) methods for multi-target tracking are essentially the same. We propose a randomized scheme, termed randomized FISST (R-FISST), where we sample the highly likely hypotheses using Markov Chain Monte Carlo (MCMC) methods which allows us to keep the problem computationally tractable. We apply the R-FISST technique to a fifty-object birth and death Space Situational Awareness (SSA) tracking and detection problem. We also compare the R-FISST technique to the Hypothesis Oriented Multiple Hypothesis Tracking (HOMHT) method using an SSA example.