Exploiting Sparse-Temporal Dynamics Via Relevance Networks for UAV Tracking

Airborne platforms' aerial viewpoint and computational power limitations pose significant challenges for tracking in complex scenarios, while existing Siamese trackers are hindered by linear context local modeling and pseudosaliency amplification. In response to this, we design an efficient sparse-temporal relevance-based tracker for UAV tracking. Specifically, regarding the process of correlation operation, we develop a selective pseudo-saliency reuptake inhibitor (SPRI) to prevent inevitable pseudo-saliency transmission and amplification in traditional CNN models caused by sequential linear feature processing. Besides, we propose a bias scale-guided dependency mining network (BSDM) to enhance interactions by leveraging multidimensional correlations within subspace interactions at each scale through a differentiated bias architecture and deformable resampling units. Finally, with a bio-visual temporal contexts integration (BTCI) module concretizing the attentive mechanism in biological vision implemented, we achieve a high-quality trade-off between tracking speed and performance by convolutionally mapping and omitting positional coding. Extensive experiments demonstrate the superiority of our proposed sparse-modeling tracker.