Online Learning-Based Hybrid Tracking Method for Unmanned Aerial Vehicles.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Outdoor tracking of unmanned aerial vehicles (UAVs) is complex due to their dynamic motion, size variations, and appearance changes.
• Existing tracking methods struggle with the challenges posed by diverse UAVs and environmental conditions.

Purpose of the Study:

• To propose an efficient hybrid tracking method for UAVs that integrates detection and tracking.
• To enhance the robustness of UAV tracking by implementing an online feature update mechanism.

Main Methods:

• Developed a hybrid tracking system combining a deep learning-based detector, a tracker, and an integrator.
• The integrator updates target features online during tracking to adapt to changes.
• Trained the detector using custom and public UAV datasets.

Main Results:

• The proposed method demonstrates effectiveness and robustness in challenging scenarios like out-of-view and low-resolution conditions.
• Experimental results on UAV123 and UAVL datasets confirm the method's generalizability.
• The system shows strong performance in UAV detection tasks.

Conclusions:

• The hybrid tracking method with online feature updates significantly improves UAV tracking performance.
• The approach effectively addresses challenges like object deformation and background changes.
• This method offers a robust solution for real-world UAV monitoring and detection.