Robust UAV Tracking via Information Synergy Fusion and Multi-Dimensional Spatial Perception

In the field of UAV tracking, existing trackers based on Siamese networks fails to fully exploit the context of the target and capture the correlations between the template and search images. These issues reduce their effectiveness in UAV tracking, especially in complex environments with challenges such as occlusion, similar object, scale variations and so on. To this end, we propose a novel robust UAV tracking method based on information synergy fusion and multi-dimensional spatial perception, named SiamFM. First, coordinate attention (CA) is introduced to embed position-aware information into the channel dimension, effectively enhancing the feature extraction capability of backbone network. Second, an information synergy fusion (ISF) module is designed to execute multiple depth-wise cross-correlation operations and accumulate the results, which gradually refines and extracts deeper feature information to accurately capture the tracking target. Then, a multi-dimensional spatial perception (MSP) module is constructed to compute global dependencies and spatial dependency information during the tracking phase, which aggregates target feature information from different scales. This new operation allows the tracker to robustly adapt to changes in target scale and proportion. Finally, comprehensive experiments conducted on traditional tracking benchmarks such as DTB70, UAV123, UAV20L, Visdrone2018, and UAVDT demonstrate that SiamFM achieves real-time performance while effectively addressing various challenging scenarios, maintaining high robustness and accuracy. For the average running speed, our tracker can reach 40 FPS, meeting the needs of UAV real-time tracking.