Enhancing region-based object tracking with the SP-SIFT feature

Existing tracking approaches are design-varied. Several features have been proposed to describe low, mid and high level cues and used to drive frame to frame tracking correspondences under mainly two strategies: purely matching and discriminative matching. However, despite the enormous amount of approaches proposed, single object tracking is still an unresolved task. In this paper we explore a, to our knowledge, new tracking approach that combines the cues at each tracking level that are currently considered the best among the existing possibilities. In this sense, point of interest (POI) descriptions are used to characterize mid-level region-based cues. These cues define a discriminative scheme whereby new instances are either classified as tracked target or surrounding background. In order to assess the advantages of our proposal, we have selected a top-ranked algorithm in the state-of-the-art, based on the use of mid-level cues, and we have integrated the SP-SIFT feature, a POI descriptor, into its scene model and decision scheme. Results indicate that the proposed approach outperforms the selected base tracking algorithm by increasing the discriminative capacity of its description feature.