Siamese Tracking Algorithm for UAVs Based on Biological Eagle-Eye Vision Mechanism

Siamese-based trackers have emerged as advanced visual object tracking algorithms for unmanned aerial vehicles (UAVs), evolving through various computational strategies and template-matching techniques. Existing Siamese trackers often lose track of target UAVs when confronted with aerial-specific issues posed by rapid relative motion, changes in scales and appearances, as well as highly-disruptive aerial environments. To address these challenges, a Siamese tracking algorithm based on biological eagle-eye vision mechanism (SiamBEM) for UAVs is proposed in this paper. Specifically, the high contrast sensitivity and adaptive target focusing module are introduced. These additions compensate for the loss of global information and restricted receptive fields in the backbone of existing trackers, particularly in scenarios involving rapid relative motion. To overcome insufficient understanding of the overall context of the target, the dual fovea visual interactive mechanism is employed to replace the traditional depth-wise cross correlation method to improve tracking stability when facing changes in scales and appearances. Furthermore, the classification loss and discrepancy loss are introduced by imitating the stimulus competition and selection mechanism. This strategy aims to improve the classification accuracy and enable better alignment of classification confidence with regression information under highly-disruptive aerial environments. Comparative experiments are conducted on six datasets, and results indicate that our method is superior to other Siamese-based methods for UAV tracking.