Multi-target tracking (MTT) refers to the problem of jointly estimating the number of targets and their trajectories with a random number of measurements, due to the births and deaths of targets. In the past several years, a number of MTT methods have been proposed, with the most popular being the joint probabilistic data association filter (JPDAF) , multiple hypothesis tracking (MHT) and random finite set (RFS) .
JPDAF and MHT as well as many traditional MTT solutions are applied by single-target filtering and data association. The single-target filter of the JPDAF and MHT uses a state-space approach based on the Bayesian framework. The most famous filter is the Kalman filter (KF), which is proposed for linear motion and Gaussian noise models . For the non-linear model, the decentralized Kalman filter (DKF) , and extended Kalman filter (EKF) , are proposed with the first order Taylor expansion and have been used to track speakers . The particle or sequential Monte Carlo (SMC) method is class of approximate numerical solutions to the Bayes recursion which can be applicable to nonlinear non-Gaussian dynamic and observation models . Compared to the KF, the PF is more robust for nonlinear models as it can approach the Bayesian optimal probability distributions of interest with a sufficiently large number of particles . It has been widely employed for speaker tracking problems , , . For example, in , independent audio and video measurements are fused for simultaneous tracking of multiple speakers. However, the mains challenges of PF are weight degeneracy and Curse-of-dimensionality. The weight degeneracy means that all but one particle will have negligible weight after a few iterations. To solve the weight degeneracy, a larger number of particles will introduce extra computational burden. For the curse-of-dimensionality, the computational cost of the approach increases dramatically as the dimensianlity of the state space increases. The data association is used to associate each measurement with an appropriate target. With these single-target tracking above, the MHT and JPDA are applied to track multi-target. MHT uses the unique-neighbour data association which associates each measurement to one of the existing tracks. The main benefit of the MHT filter is maintaining multiple hypotheses of the association between target states and the measurements. However, the number of hypotheses grows exponentially over time . JPDA uses all-neighbours data association which uses all the measurements for updating the entire track estimate.