Global Image Correlation Filter with H-D Fusion Mechanism for Visual Tracking

The discriminative correlation filter has remarkable robustness under various circumstances in the visual tracking of videos. Nevertheless, there are some challenges for the further improvement of the tracking performance, including confusion between the tracked target, complex background, and underfitting because of a lack of training samples. The main contribution of the study is developing a new framework which combines with global image sampling enable to significantly augment the negative samples without the need of corruptting those positive samples. Besides we propose a fusion mechanism by combining two image patch representations in accordance with their confidence scores. Obtained results show that the proposed method is an effective scheme in leveraging the complementary properties of deep and hand-crafted features. Based on the conducted experiments on several benchmarks such as OTB2013, TB-50, and TB-100 datasets, it is concluded that the proposed approaches can achieve better result compared to the state-ofthe-art methods.