Appearance features for online multiple camera multiple target tracking
Résumé
Multiple object tracking methods in the state-of-the-art are challenged by appearance variation, environment changes and longterm occlusions. Exploiting multiple calibrated and frame synchronized cameras holds the promise of alleviating these problems, in particular, the one pertaining to occlusion. The practical realization of this idea faces the problem that the appearance of the same target can change through different cameras. Thus, particular care should be taken in order to enhance the computation of appearance distances between targets in multiple cameras. In this paper, we tackle the problem of multiple object multiple camera tracking by adopting a Markov Decision Process framework. We concentrate on the effect of the affinity function by discussing different possible implementations and validating their performance, in terms of the MOT metric and the ID measure, on the PETS 2009 and EPFL datasets. Our experimental result shows a significant improvement of multiple cameras approaches with a sufficiently large overlapping zone compared to single camera ones.
Origine : Fichiers produits par l'(les) auteur(s)