Labeled Multi-object Tracking Algorithms for Generic Observation Model

In this paper, we are devoted to the labeled multiobject tracking problem for generic observation model (GOM) in the framework of Finite set statistics. Firstly, we derive a product-labeled multi-object (P-LMO) filter following by the PLMO formed density which is a more detailed expression for the general labeled multi-object density [1]. The proposed PLMO filter is a closed form solution to labeled multi-object Bayesian filter under the standard multi-object transition kernel and generic multi-object likelihood and thus can be used as the performance benchmark in labeled multi-object tracking. Secondly, we propose a generalization of LMB filter, named LMB filter for GOM by approximating the full multi-object density as a class of LMB density matching the original labeled first order moment as well as minimizing the Kullback-Leibler divergence from the original multi-object density. The LMB-GOM filter can be seen as a principled approximation of P-LMO filter, which not only inherits the advantages of the multi-Bernoulli filter for image data with the intuitive mathematical structure of multi-Bernoulli RFS, but also the accuracy of P-LMO filter with less computation burden. In numerical experiments, the performance of the proposed algorithms are verified in typical tracking scenarios.