Learning Response-Consistent and Background-Suppressed Correlation Filters for Real-Time UAV Tracking.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Discriminative Correlation Filter (DCF) methods are prevalent in unmanned aerial vehicle (UAV) target tracking due to their efficiency.
• UAV tracking faces challenges like clutter, occlusion, and fast motion, leading to response map interferences and target loss.

Purpose of the Study:

• To develop a robust correlation filter for UAV target tracking that addresses challenges causing target drift and loss.
• To improve tracking accuracy and reliability in complex UAV operational environments.

Main Methods:

• A response-consistent module was developed, generating and aligning response maps from adjacent frames using an L2-norm constraint.
• A background-suppressed module was introduced, employing an attention mask matrix to reduce interference from background distractors.
• The proposed modules were integrated into the DCF framework for enhanced UAV tracking.

Main Results:

• The proposed tracker demonstrated superior performance against 22 state-of-the-art trackers on challenging UAV benchmarks (UAV123@10fps, DTB70, UAVDT).
• The method achieved real-time tracking at approximately 36 FPS on a single CPU.
• The response-consistent module prevented sudden response changes and preserved filter discriminative ability, while the background-suppressed module reduced distractor interference.

Conclusions:

• The developed response-consistent and background-suppressed correlation filter significantly enhances UAV target tracking performance.
• The proposed method offers a robust and efficient solution for real-time target tracking in complex UAV scenarios.