Ukrainian firm develops AI drone detection software for frontline use

A Ukrainian software company has developed an AI-powered drone detection system that runs entirely offline on standard consumer hardware, detects targets as small as four pixels at distances up to 2.5 kilometers depending on optics, and automatically alerts military units through messengers with photographs and coordinates within milliseconds of identifying a threat.

The system, called I-SEE, is designed to address a specific and deadly gap in frontline air defense: the seconds between when a drone first appears in camera footage and when a human operator notices it and reacts. Those seconds have determined survival outcomes for thousands of Ukrainian soldiers since Russia began its mass FPV drone campaign. An experienced fighter can maintain focus for hours but cannot sustain perfect attentiveness indefinitely, and a single lapse is enough for a fiber-optic FPV drone to close the distance and strike before any countermeasure can engage. I-SEE replaces that human attention with continuous machine processing running 24 hours a day at 30 to 60 frames per second, according to the company’s technical documentation.

Electronic warfare systems cannot reliably detect fiber-optic FPV drones or drones using signal relay systems because those platforms transmit no radio frequency signal during their attack run, becoming visible to RF-based detection only in the final phase of approach when intercept windows are already severely compressed. Radar systems partially compensate but carry costs and deployment complexity that put them beyond the reach of most frontline units.

I-SEE’s approach is optical: it uses computer vision and AI to find and track drone-sized objects in camera footage, detecting threats before a human eye would register them as anything other than background clutter.

Detection performance figures from field testing, as reported by the company, show FPV drone detection at 250 to 300 meters using a camera with two times zoom, approximately one kilometer using a five times zoom camera, and up to 2.5 kilometers using a Mavic 3T thermal camera at maximum zoom. A standard street camera with 30 times optical zoom detected drones at 1,650 meters in testing. The company notes explicitly that these figures depend on optics, zoom level, and observation conditions — a meaningful caveat in a system whose effectiveness is fundamentally constrained by the quality of the camera feeding it, not by the software itself.

The system processes video locally at the edge, running entirely offline without internet connectivity, which eliminates both the latency of cloud-based processing and the vulnerability of systems that depend on network access in electronically contested environments, per the company’s technical description. It accepts video feeds via USB, RTSP, RTMP, HTTP, and other standard protocols, meaning it integrates with cameras already installed rather than requiring proprietary hardware. The minimum hardware requirement is a standard PC or laptop equipped with a GPU. The company describes the architecture as running standard computational platforms rather than expensive custom hardware, which is the foundation of the cost advantage it claims over radar-based and hardware-centric alternatives.

The alert architecture is built for the operational reality of small units without dedicated signal officers. When I-SEE detects a drone, it automatically pushes an alert to Telegram or Discord with a photograph and the target’s coordinates, azimuth, and estimated bearing — delivering actionable intelligence to fighters’ phones in the same applications they already use for communication. The system tracks trajectory and calculates an intercept point for engagement, providing guidance for counter-drone weapons rather than simply triggering an alarm. It can simultaneously track up to 100 targets depending on computing hardware capacity, and it maintains track continuity through brief target disappearances rather than losing the contact and restarting detection from zero.

The company was direct about the product’s ambitions beyond its current capabilities. “We want to promote our development so that other hardware developers can see it and we can integrate with them,” the company stated. “Because our system can make simple devices smart and autonomous.”

That integration roadmap, as described in the technical documentation, includes completed integration with net guns and turrets for automatic engagement, Android mobile application support currently in testing, and planned future integration with electronic warfare systems, autonomous engagement capability, audio analysis, and radar inputs. The modular software architecture is designed specifically to accept new sensor types and engagement systems through API and webhook connections without replacing the core detection platform.

The false alarm problem that has undermined many visual detection systems in operational use is addressed through what the company describes as AI filtering of birds, debris, and reflections. The field testing data shows detection of targets as small as four pixels, which is the minimum signature an FPV drone presents at operational engagement distances. At that threshold, distinguishing a drone from a bird or a piece of airborne debris requires a classification model trained on real combat data rather than laboratory examples, and the company notes that regular retraining on battlefield data is required to maintain accuracy as adversary drone types evolve.

Fiber-optic FPV drones are among the most difficult targets in the current Ukrainian battlefield environment precisely because existing detection architectures were not designed for them. Software that processes camera footage already present on most military positions, runs without internet access, and sends smartphone alerts with coordinates and intercept guidance addresses that specific threat more directly than any expensive radar system that most units will never receive.