UAV target tracking method based on global feature interaction and anchor-frame-free perceptual feature modulation

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Unmanned Aerial Vehicle (UAV) target tracking relies on real-time video analysis.
• Deep learning methods, particularly Siamese-based, show promise but struggle with speed-accuracy trade-offs.
• Existing methods face challenges in refining features and handling target deformation.

Purpose of the Study:

• To enhance the accuracy and efficiency of UAV target tracking.
• To improve feature representation and reduce computational load for real-time applications.
• To boost tracking performance for small and deforming targets.

Main Methods:

• Implemented feature fusion within deep inter-correlation operations.
• Introduced a global attention mechanism to expand the field of view and refine features.
• Designed an anchor-free frame-aware feature modulation mechanism for efficient anchor generation and target refinement.

Main Results:

• The proposed algorithm demonstrates a balance between speed and accuracy on UAV tracking datasets (UAV123@10fps, UAV20L, DTB70).
• Achieved a real-time processing speed of 30 frames per second on the Jetson Orin Nano platform.
• Showcased improved tracking performance for small targets and adaptability to target deformation.

Conclusions:

• The developed method offers a robust solution for real-time UAV target tracking.
• The integration of feature fusion and attention mechanisms significantly enhances tracking capabilities.
• The algorithm's efficiency and accuracy are validated through extensive experiments and a physical platform implementation.