Online Multi-Object Tracking via Labeled Random Finite Set with Appearance Learning

In this paper, a novel approach to online multi-object tracking is proposed via Labeled Random Finite Sets (RFS) combined with appearance learning. The Labeled RFS formulation of the multi-object state naturally accommodates a time-varying number of objects, track labels, and false positive rejection in a single Bayesian framework. The proposed algorithm exploits appearance feature information for the purpose of learning an object's appearance model, and uses this additional information in the construction an augmented likelihood which improves performance and facilitates track re-initialization. This approach enhances the baseline tracking algorithm and shows better performance with respect to mis-detections, occlusions and false track rejection. Competitive tracking results are shown compared to state-of-the-art algorithms on PETS benchmark [1] video datasets.