An object tracking framework with recapture based on correlation filters and Siamese networks

Recently, both correlation filters-based and Siamese-based trackers have achieved great progress in the field of visual object tracking. Whereas, some trackers perform not that well under some tough situations. Aimed at occlusions and background clutters, we propose a tracking framework combining correlation filter tracking and Siamese-based object tracking. First, we combine deep features with handcrafted features in correlation filter learning. Then, we propose a new robust criterion to evaluate the robustness of the tracking results and decide whether to start the Siamese tracking model according to the criterion. If the robustness evaluation value is lower than the adaptive threshold, we start the Siamese tracking for object recapture. We conduct the experiment to evaluate our tracker in four datasets to validate the effectiveness. The experimental results show that our method achieves state-of-the-art performance. It is worth mentioning that our tracker performs favorably against many existing trackers in the TC128 dataset.

Automated summary

This study proposes a new framework that combines correlation filter tracking and Siamese-based object tracking to address the performance issues of trackers in occlusions and background clutter. Experimental results demonstrate the effectiveness of our method across multiple datasets, particularly outperforming existing trackers on the TC128 dataset.