Benchmarking Deep Trackers on Aerial Videos

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Deep learning visual object trackers excel on ground-level benchmarks.
• Aerial tracking introduces unique challenges not present in ground-level scenarios.

Purpose of the Study:

• To evaluate the performance of ten deep learning-based visual object trackers on aerial datasets.
• To compare different tracking approaches including tracking by detection, discriminative correlation filters, Siamese networks, and reinforcement learning.

Main Methods:

• Experiments were conducted on four aerial datasets: OTB2015 (aerial subset), UAV123, UAV20L, and DTB70.
• Ten state-of-the-art deep learning trackers employing diverse methodologies were selected for comparison.

Main Results:

• All evaluated trackers demonstrated significantly reduced performance on aerial datasets compared to ground-level videos.
• Key challenges identified include smaller target size, complex camera motion, target rotation, out-of-view movement, and environmental clutter.

Conclusions:

• Current deep learning trackers are not optimized for the complexities of aerial visual object tracking.
• Further research is needed to develop robust trackers capable of handling aerial surveillance and navigation challenges.