Siamese object tracking based on multi-frequency enhancement feature

Tracker based on the Siamese network is an important research direction of target tracking. However, Siamese trackers have the problem of insufficient characteristic expression ability extensively. To improve the situation, we introduced the effective attention module multi-frequency attention, which considers the attention mechanism from the frequency domain perspective and effectively raises the efficiency of feature representation of the tracker. Meanwhile, densely connected neural networks are applied to the tracker, which integrates the surface orientation information and deep-seated semantic features about the object, which will contribute to enhancing the localization and expression ability of the tracker. The proposed method achieves excellent results on Siamese network tracking. In order to prove the effectiveness of our strategy, we conducted experiments on benchmarks, including OTB100, OTB2013, VOT2016, and VOT2018, and the obtained results indicate that strategy we proposed achieves high accuracy and efficiency and still keeps outstanding with the effect of occlusion and illumination. Experiments show that the proposed method fulfills real-time tracking.