A new computer-vision method turns scattered event-camera signals into continuous motion trails to catch distant, tiny drones that conventional detectors miss.
Event cameras record pixel-level brightness changes instead of full frames, which makes them fast and power-efficient for spotting fast-moving objects. But a drone far from the camera produces only a handful of scattered events, easily lost in background clutter or the camera's own motion. Researchers describe PointEvent, a lightweight model that treats those scattered events as pieces of an ordered path rather than isolated points. It sorts the same event stream two ways, by location and by time, and scans both orders together to stitch fragmented motion evidence back into a coherent trail, while a high-resolution branch preserves fine detail and a separate module filters out clutter.
Tiny-object detection is the weak link in drone-spotting systems, whether for airspace safety, wildlife surveys, or counter-UAV security, and most previous gains have come from throwing more compute at the problem. PointEvent's pitch is the opposite: better organization of the same sparse data, in a model the paper claims has fewer parameters and faster inference than the methods it's compared against.
The claims rest on the authors' own benchmarks, not independent testing, so treat them as promising rather than proven until other labs run PointEvent against their own footage.