Abstract:

In single-object tracking， the accuracy of the tracking bounding box is often compromised by factors such as deformation， motion blur， occlusion， and background interference. In particular， background interference frequently leads to tracking hopping and drift. To mitigate these issues， a two-stage tracking algorithm that integrated motion information with a dual-attention mechanism was proposed. In the first stage， a SiamCAR tracker with a dual-attention mechanism was employed to coarsely locate the target in the current frame. In the second stage， a refinement module of the bounding box was constructed using pixel-level similarity computations to learn the subtle features of the target under low-latency conditions， thereby enhancing the tracking accuracy. Finally， the tracking box obtained based on appearance features was fused with the target’s motion trajectory information to mitigate tracking drift and hopping. Experimental results on the OTB100 dataset indicate that the success rate and accuracy of the tracking box have improved by 4.6% and 2.8%， respectively， compared to the original. The success rate in the presence of background interference has reached 69.6%.

Key words: single object tracking, SiamCAR, Siamese network, neural network, attention mechanism