EventRadar uses event cameras to detect UAVs from over 1 km away

EventRadar, developed by Zhiting Zhou and colleagues from Tsinghua University, presents a novel approach to long-range unmanned aerial vehicle (UAV) detection. Traditional computer vision methods rely on spatial cues like silhouette or track continuity, which degrade at distances beyond 700 meters. EventRadar instead leverages propeller-induced temporal periodicity captured by event cameras—an emerging sensing modality that records pixel-level brightness changes asynchronously. The system combines two key innovations: Scene-Anchored Geometry Evidence (SAGE) which fuses event streams with IMU orientation data to maintain a bearing-indexed scene memory, filtering out persistent clutter; and a Comb-guided Harmonic-Group Learned Iterative Shrinkage and Thresholding Algorithm (CHG) that treats each candidate as a weak high-rate timing signal, recovering phase-insensitive harmonic evidence with fixed compute.

On real-world UAV recordings from 700 to 1500 meters, EventRadar achieves 0.990 mean average precision (mAP@0.3) and 0.949 F1 score, reducing false negatives to just 0.009—significantly outperforming previous event-camera baselines. The prototype also demonstrates real-time feasibility, making it practical for protecting airports and sensitive sites from unauthorized drone incursions. By shifting the detection paradigm from appearance to motion periodicity, EventRadar opens a new path for visual sensing at kilometer scales where traditional sensors fail.