Combining complementary trackers for enhanced long-term visual object tracking

Abstract

Several different algorithms have been studied to combine the capabilities of baseline trackers in the context of short-term visual object tracking. Despite such an extended interest, the long-term setting has not been taken into consideration by previous studies. In this paper, we explicitly consider long-term tracking scenarios and provide a framework to fuse the characteristics of complementary state-of-the-art trackers to achieve enhanced tracking performance. Our strategy perceives whether the two trackers are following the target object through an online learned deep verification model. Such a target recognition strategy enables the activation of a decision strategy which selects the best performing tracker as well as it corrects their performance when failing. The proposed solution is studied extensively and the comparison with several other approaches reveals that it beats the state-of-the-art on the long-term visual tracking benchmarks LTB-35, LTB-50, TLP, and LaSOT.