SkyEV: RGB-Event UAV detection and tracking dataset and baseline

Detecting Unmanned Aerial Vehicles (UAVs) is a critical security priority, yet current vision-based detection systems struggle with the fundamental trade-off between Field of View (FOV) and detection range. While traditional RGB cameras provide rich semantic information, they are often limited by fixed shutter times and motion blur. Event-based cameras offer high temporal resolution and dynamic range, but they lack the spatial density required for precise identification. Existing datasets for UAV detection often fail to reflect realistic operational conditions, such as camera ego-motion, uncompressed image quality, and the extreme small-scale nature of distant targets.

The authors developed SkyEV, a dataset designed to bridge the gap between idealized laboratory conditions and real-world counter-UAV scenarios. The setup utilizes two event cameras and one global-shutter RGB camera mounted in parallel. By hardware-timestamping the RGB shutter signal within the event stream, the researchers achieved precise temporal synchronization. The dataset includes 398 sequences, totaling over 2.17 hours of footage, featuring eight different drone types. To facilitate research, the authors provide a unified data loader and a baseline multi-modal detection architecture that fuses RGB frames with event-based histograms using a modified Scene Adaptive Sparse Transformer (SAST) and a YOLOX backbone.

SkyEV addresses the scarcity of high-quality, multi-modal data for small-object detection. By including significant camera motion and diverse optical setups, it forces models to handle background clutter and varying illumination—common challenges in real-world security applications. The inclusion of uncompressed 10-bit raw RGB frames and high-speed event data allows researchers to explore advanced sensor fusion techniques that are essential for detecting fast-moving, small-scale drones at long ranges, where traditional frame-based systems often fail.