Real-time object tracking via self-adaptive appearance modeling

Abstract

One of the main factors that limit the accuracy and robustness of visual tracking algorithms is the lack of suitable appearance models. The robustness and effectiveness of object appearance models is severely affected by the changing object appearances during the tracking process and the interference of other similar objects around the truth object. In this paper, a self-adaptive appearance model pool based on multi-sample is constructed to improve the robustness of the object appearance models. In order to deal with variable object states, the initial sample given in the first frame and the samples generated in the subsequent tracking process are combined into a sample set to represent various appearances of the object. In addition, a dynamic selection strategy is explored to update and maintain the sample components that are derived from varieties of sources. In order to distinguish the tracking object from other similar candidate objects, multi-feature response fusion strategy is proposed, which can effectively improve the expression ability of the appearance model. Extensive experiments on the popular benchmark datasets demonstrate that the proposed tracking approach performs favorably against several other state-of-the-art tracking algorithms.