Multi-class Multi-target Tracking with the Poisson Labeled Multi-Bernoulli filter

Abstract

A Random Finite Set (RFS) based multi-target multi-class filter is proposed, which utilizes a labeled Multi-Bernoulli distribution to model the multi-target state, together with a Poisson RFS distribution to model potential new targets. The Poisson distribution is advantageous in modelling any number of new targets with different classes. This is because it allows birth targets with different expected numbers per class to be modelled, where each class of targets could be modelled to appear in different locations of the map. Using class information is expected to improve overall tracking performance, when sensors provide class information, such as video camera image classifiers. The proposed filter is referred to as the Class Poisson Labeled Multi-Bernoulli (CPLMB) filter. Results show that augmenting the target state with class, by using classification information in the measurements, increases PLMB filter performance and, after few iterations, each target class converges to a single value with probability close to or equal to unity.