Multiclass Generalized Labeled Multi-Bernoulli Filter With Kernel Density Estimation for Target Tracking in Unknown Backgrounds

In multitarget tracking (MTT), the prior information of background parameters, such as clutter intensity and detection probability, exerts a significant influence on the performance of the filter. Thus, when performing MTT in unknown backgrounds, the filter should possess the capability to estimate background parameters. To achieve the online estimation of the unknown background parameters, this article introduces a multiclass generalized labeled multi-Bernoulli filter based on the kernel density estimation (KDE-MC-GLMB). The proposed filter uses MC-GLMB to model the clutter generators and targets of interest, respectively. Via propagating the joint density forward in time, the filter can jointly estimate the clutter intensity and multitarget states. To adapt to the unknown clutter spatial density, a KDE-based clutter estimator is incorporated into the MC-GLMB recursion, thereby enabling accurate estimation of the clutter spatial density. In addition, to enhance computational efficiency, we implement L-scan truncation on the number of Gaussian kernels in KDE and introduce a clutter density updating strategy based on the best association map (BAM) to limit the repetition of KDE processing. As for the unknown detection probability, the Beta-Gaussian mixture implementation of the proposed filter is provided in this article as well. The simulation results demonstrate that, in the unknown background environment, the KDE-MC-GLMB filter exhibits superior tracking performance compared to existing random-finite-set-based robust filters.