Correlation Filters Based on Multi-Expert and Game Theory for Visual Object Tracking

Most existing correlation filter-based tracking algorithms only use the linear additive features or simple fusion features, which are unable to fully learn the appearance information of the target. In this article, a multi-expert and game theory correlation filter (MEGTCF) algorithm is proposed to treat the aforementioned problems. The algorithm extracts a variety of handcrafted features and deep features of the target and forms multiple experts in different combinations to track the target in parallel. In addition, the game theory strategy is introduced to obtain more reliable decisions. The expert evaluation mechanism evaluates and selects the most discriminative experts as game participants and fully explores the complementary information among different expert features through the iterative game, so as to realize the information interaction between expert features. We also propose a confidence indicator to calculate the weights of the experts with higher scores and adaptively update the expert model to avoid expert corruption. Extensive experiments are conducted on five popular benchmarks, including OTB2013, OTB2015, VOT2018, LaSOT, and the UAV123 tracking benchmark. The success rate and precision of the MEGTCF tracker on the OTB2015 reach 67.8% and 91.4%, respectively, which outperforms other state-of-the-art trackers based on correlation filters. In addition, the MEGTCF tracker is also highly suitable for challenging scenarios such as object deformation, background clutter, and occlusion.