Track Classification for Random Finite Set Based Multi-Sensor Multi-Object Tracking

The state-of-the-art of random finite set (RFS) based approaches for multi-sensor multi-object setups solve the classification and track estimation jointly in a Bayesian style. This is computationally demanding and often requires additional modeling and parameter estimation. Additionally, these approaches are not designed to make use of direct class estimations, e.g., from machine learning detectors, but estimate the class based only on the kinematic features. This work applies a separated track classification, which uses direct class estimations, to RFS-based trackers. The proposed approach can be implemented for various RFS-based multi-sensor multi-object tracking algorithms without altering their structure and without additional modeling effort. For the update of the class estimation of a track, three different methods are presented. The three approaches are demonstrated and evaluated in combination with a labeled multi-Bernoulli filter on simulated and real-world data.