MLT-STARK

Memory-Augmented Long-Term Anti-UAV Tracking

MLT-STARK Architecture

Abstract

Unmanned Aerial Vehicles (UAVs) are extensively utilized in geoscience and remote sensing for tasks ranging from crop monitoring to security surveillance. Their widespread use, however, raises concerns about safety and privacy, especially in sensitive areas like military zones. Tracking UAVs is crucial to mitigate risks of misuse and collisions with manned aircraft, yet existing methods struggle with challenges such as recurrent target disappearance and sudden camera movements in Anti- UAV scenarios. Tracking UAVs is much more challenging in TIR videos due to the small size of targets and significant variations in their appearance. To tackle these challenges, we introduce MLT-STARK (Memory-based Long-Term STARK) for Anti-UAV Visual Tracking, which combines a memory-based dynamic detection module with STARK, a Transformer-based tracker. The memory-based dynamic detection module lever- ages historical tracking cues to improve target re-identification, enabling adaptive responses to occlusions and sudden motion shifts. This integration enhances the tracker’s capability to detect and recover lost targets in challenging tracking environments. When evaluated on the IR Anti-UAV and visible DUT datasets, LT-STARK demonstrated significant performance enhancements over baseline models, achieving a 6.48% increase in success rate and a 9.54% boost in precision on the Anti-UAV Dataset, and a 32.68% higher success rate with a 26.22% improvement in precision on the DUT Dataset. These results demonstrate the effectiveness of the LT-STARK model in addressing the chal- lenges associated with Anti-UAV tracking, marking a significant contribution to UAV intrusion detection and prevention

Novelty

  • Improved object localization: By integrating a detector into the tracker, the model can encode spatial information to relocate and track objects more effectively.

  • Enhanced object identification: The proposed dynamic detector module provides a valuable feature by offering object flags and enabling more efficient re-identification of lost objects within the current state-of-the-art (SotA) Transformer tracker framework.

  • Memory-based detection for robust tracking: By introducing a historical memory mechanism, the proposed Memory-Based Dynamic Detection Module (M-DDM) significantly enhances tracking robustness by leveraging past templates for target re-identification.

Results

References