AE-UAV: An Air-to-Air Event-Based UAV Tracking Benchmark and a Real-Time Frequency-Domain Tracker

Research Question and Approach

Air-to-air (A2A) tracking is critical for airborne remote sensing, yet traditional frame-based cameras fail in these scenarios due to motion blur, limited dynamic range, and fixed temporal sampling. While event cameras offer microsecond resolution and high dynamic range, research has been hindered by a lack of specialized A2A datasets and the high computational cost of existing trackers, which typically require GPU acceleration. To address these gaps, the authors introduce the AE-UAV benchmark and the Fast-Slow Frequency-domain Tracking (FSFT) method.

The AE-UAV Benchmark

AE-UAV is the first event-based dataset captured from an aerial platform specifically for A2A tracking. It includes 178 flight sequences featuring diverse illumination conditions (daylight, backlit, and night) and complex relative motion patterns (pursuit, evasion, and head-on approaches). A key innovation is the use of a continuous-time annotation framework based on cubic B-spline interpolation. Unlike traditional methods that assign constant bounding boxes between keyframes, this approach generates C2-continuous trajectories, allowing for consistent evaluation at arbitrary temporal resolutions.

The FSFT Tracker

FSFT is a training-free, lightweight framework designed for resource-constrained UAV hardware. It operates on a dual-pathway architecture:

• Fast Pathway: Performs per-packet frequency-domain localization using Kalman prediction and template matching to track targets at high speeds.
• Slow Pathway: Executes periodic detection-based drift correction to maintain accuracy over time.

By processing event packets directly and avoiding frame reconstruction, FSFT achieves 420 FPS on CPU-only hardware. Experimental results show that it retains 93.97% of the accuracy of state-of-the-art GPU-dependent methods while providing a 5.32-fold speedup and superior generalization across different temporal resolutions.