MTV: Selective State Space Tracker for Infrared UAV

Infrared unmanned aerial vehicle (UAV) tracking is critical for low-altitude security systems, providing robust all-weather capabilities for detecting and predicting the trajectories of unauthorized aircraft. Unlike visible-light systems, infrared tracking leverages lighting-invariant thermal signatures, enabling reliable day-and-night operation. However, infrared tracking faces significant challenges: limited texture information, high target-background similarity in complex environments, and tracking drift during long-term monitoring of small, fast-moving UAVs. Conventional trackers relying on shallow features struggle to adapt to the dynamic motion and small size of infrared UAVs. In contrast, deep learning-based Siamese trackers leverage high-level semantic features, balancing real-time performance and accuracy. However, their reliance on linear cross-correlation limits their ability to model complex non-linear relationships, while transformer-based trackers are computationally intensive and susceptible to background interference. To address these limitations, we propose MTV, a selective state-space tracker that integrates a Region Routing Network (RRN) for refined feature correlation and a Temporal-Spatial Template Updater (TSTUpdater) to mitigate temporal inconsistencies, achieving robust long-term tracking. Extensive experiments on the Anti-UAV410 dataset demonstrate state-of-the-art performance, achieving a precision of 68.44 % using only ImageNet pretraining and establishing a new baseline for infrared UAV tracking.