Multi-Object Tracking with Data Association by a Similarity Identification Model

Abstract

This paper presents a novel approach for Multi-Object Tracking by a visual-based identification of the objects. Unlike systems using the learning of an appearance model for every tracked object, the proposed data association method is based on only one similarity model capable of discriminating whether a detected person in the current frame corresponds, or not, to one of the tracked ones. This method allows an online tracking of any target, selected by a user over a live sequence, without requiring previous knowledge of it. The experimental results show the excellent performance of the discriminative task realized by our similarity identification model. Furthermore, this model has been integrated into a whole Multi-Object Tracking algorithm, which has been evaluated over challenging public datasets with successful results. The comparison of these results with other state-of-the-art tracking-by-detection approaches has proved the remarkable improvement obtained by the method proposed.