A drone that costs less than a laptop can shut down an international airport. In the autumn of 2025, Europe learned this the hard way. Over a span of weeks, unexplained drones forced repeated closures at Copenhagen and Oslo, Munich, and a string of other airports and military bases, grounding flights, diverting aircraft, stranding thousands of passengers, and triggering cross-border scrambles. In one Munich incident alone, seventeen flights were grounded, and roughly 3,000 passengers were affected. No payload was dropped. No shot was fired. The mere presence of an aircraft nobody could identify was enough.
That asymmetry, a thousand-dollar machine imposing million-dollar consequences, is the single most important fact in security today. It explains why counter-unmanned aircraft systems (C-UAS) have moved, in the space of a few years, from a niche defense curiosity to a board-level priority for airports, utilities, ports, prisons, stadiums, and governments. And it explains why the most interesting question in the field is no longer whether to defend the airspace, but how to see what's in it, reliably, autonomously, and at scale.
This article makes the case for a specific answer to that question: an AI "brain" that detects, classifies, and tracks aerial threats using visual data and computer vision alone.
Part 1: Why we need counter-UAS, and why now
The drone threat is no longer hypothetical, and it is no longer rare. A few data points sketch the shape of it:
- In the United States, the FAA recorded 411 illegal drone incursions near U.S. airports in the first quarter of 2025 alone, and the agency says it still receives more than 100 drone sightings near airports every month.
- In 2024, over 350 drone detections were recorded across 100 different U.S. military installations, according to the head of NORAD and U.S. Northern Command.
- Drones have been used against critical infrastructure for years, including a 2020 attempt to short-circuit a Pennsylvania electrical substation with a drone trailing copper wire, which federal agencies called the first known case of a modified drone targeting U.S. energy infrastructure.
- Organized crime has industrialized drone smuggling. In Georgia alone, authorities arrested around 150 people in a multi-state enterprise using drones to deliver drugs, phones, and weapons into prisons.
Layered on top of this civilian picture is the battlefield. In Ukraine, small, inexpensive drones have become central to modern warfare, and Western and Ukrainian officials estimate that drones now cause roughly 70–80% of battlefield casualties, though those figures come largely from official statements rather than independently verified data. That conflict has become a live laboratory for both drone attack and drone defense, and the lessons are propagating directly into the procurement plans of every Western military.
The economics are what make this a durable problem rather than a passing one. Traditional security spending is justified by probability: how likely is an attack? Drone risk flips that logic to consequence: how expensive is even one interruption? When the barrier to attempt is a hobby-grade quadcopter, and the cost of response is a full airport shutdown, a perimeter sweep, and a public record of vulnerability, the attacker wins even when nothing is actually destroyed. Counter-drone capability, in other words, is becoming less like a one-off hardware purchase and more like cybersecurity monitoring, a permanent layer of operational resilience.
Three forces are converging to turn that recognition into a market.
The threat is escalating and visible. The 2025 European airport wave was widely read as adversaries probing Western defenses with minimal effort and deniability; Danish authorities publicly treated the Copenhagen incident as a possible hybrid attack on critical infrastructure. That is exactly the kind of asymmetric pressure that forces governments to act.
The regulation is finally catching up. For years, the binding constraint in civilian airspace wasn't technology, it was the authority. It was simply illegal for most operators to interfere with an aircraft, even a hostile one. That is changing. In the U.S., the SAFER SKIES Act was folded into the FY26 National Defense Authorization Act and signed into law on December 18, 2025, giving trained and certified state, local, tribal, and territorial law enforcement and correctional agencies a path to lawfully detect, track, and mitigate drones that pose a credible threat to people, critical infrastructure, major events, or prisons. In parallel, the FAA published a long-awaited proposed rule in May 2026 that would let critical-infrastructure facilities petition for drone flight restrictions over their sites. And ahead of the 2026 FIFA World Cup, FEMA awarded an initial $250 million in counter-UAS grants to the host states and the National Capital Region.
The money has arrived. Forecasts vary by methodology, but they all point the same direction, sharp, sustained growth. Estimates of 2026 C-UAS spending range from roughly $4 billion to more than $14 billion, with compound annual growth rates clustered between roughly 18% and 26.5% over the coming years. The signal in the noise is the contract flow: in March 2026, the U.S. Army awarded Anduril a ten-year enterprise contract with a ceiling of up to $20 billion, with a counter-drone command-and-control order as its first task order. The first quarter of 2026 alone reportedly saw more than $29 billion in publicly announced C-UAS contracts worldwide.
A growing threat, a clear legal path, and a flood of capital. That is why now.
Part 2: The detection problem nobody talks about
Most public conversation about counter-drone systems fixates on the dramatic end of the "kill chain", the jammers, the nets, the lasers, the kinetic interceptors. But you cannot neutralize what you cannot reliably see, and detection is where the field actually struggles. There are four classical ways to detect a drone, and each one breaks in a way that matters.
Radio frequency (RF) sensors listen for the communication link between a drone and its operator. They're passive, cheap, and can sometimes fingerprint a drone down to its model. But they share a fatal assumption: that the drone is emitting at all. RF interception fails against autonomous platforms that radiate little.
Radar detects objects rather than signals, so it sees a silent drone the same as any other moving thing. Its problem is discrimination. The radar cross-section of a small consumer quadcopter is often similar to that of a bird, which drives up the false-alarm rate. In dense urban environments, clutter and low-altitude flight near terrain can blind radar to low-flying threats entirely.
Acoustic sensors recognize the sound signature of propellers and motors. They're inexpensive and combat-proven at scale in Ukraine, but their range is short and they degrade in noisy environments, and the newest, quietest drones are designed to evade them.
Electro-optical and infrared (EO/IR) cameras provide visual confirmation, the one thing the other modalities can't deliver. In the classical architecture, however, the camera was relegated to a supporting role: a pan-tilt unit that another sensor cued to a probable location, after which a human operator squinted at the feed to decide whether it was a threat. EO was the witness, not the detective.
The deeper problem is false alarms. This is one of the most underreported realities of fielded C-UAS: a system tuned aggressively enough to catch every small drone will generate so many nuisance alerts including birds, weather, insects, reflections, that operators stop trusting it. Alarm fatigue sets in, and real threats start getting ignored. A counter-drone system that cries wolf is worse than no system at all, because it manufactures the very complacency an attacker needs. (I went deeper on this in Bird or Drone? The Hardest Classification Problem in Counter-UAS.)
Part 3: The spectrum is going dark
The strongest argument for a visual-first approach isn't a weakness of the alternatives. It's a trend: the threats are evolving in a direction that makes the non-visual modalities obsolete.
The entire premise of RF detection and RF jamming is that a drone is tethered to the electromagnetic spectrum. That premise is collapsing.
- Fully autonomous drones navigate by inertial measurement, optical flow, and onboard vision. They take off, fly a mission, and return without ever talking to an operator. Pre-programmed missions need no link at all, which means there is no command-and-control signal to detect and no link to jam.
- Fiber-optic tethered drones, now widespread on both sides in Ukraine, run a literal physical hair-thin cable between operator and aircraft. They are completely decoupled from the spectrum, invisible to RF detectors and immune to RF jammers by design.
- GPS-denied and vision-guided navigation means a drone can fly precisely through an environment where satellite navigation is unavailable or actively spoofed, removing yet another signal an adversary might exploit or a defender might detect.
Put plainly: the next generation of threats will be RF-silent. A defense architecture built primarily on listening for radio emissions is preparing to fight the last war. Against a dark, autonomous drone, you are left with exactly two physical phenomena that cannot be switched off and the object reflects light, and the object has a shape that moves through space. Both of those are the native domain of computer vision.
Part 4: Why vision-only, and why it had to wait for modern AI
A vision-first detection brain has a set of structural advantages that look increasingly decisive as the threat evolves.
It's passive. A camera emits nothing. It doesn't announce its presence, doesn't interfere with the radio spectrum, doesn't require a transmission license, and can't be detected by the drone it's watching. For protecting sensitive sites, covert and emission-free is a feature, not a footnote.
It answers the only question that matters legally. Detection tells you something is there. Classification tells you what. And the visual layer is the one that can distinguish a weaponized threat from a hobbyist, a delivery drone, or an authorized survey platform. You don't jam a police drone. Under the new mitigation authorities, the legal trigger is a "credible threat", and establishing credibility is fundamentally a perception-and-classification problem, not a signal-detection one.
It solves the bird problem at the right level. Radar confuses a quadcopter with an osprey because, to radar, they're similar-sized blobs. A vision model doesn't reason about radar cross-section. Instead, it reasons about appearance, structure, and motion. A bird flaps; a quadcopter holds rigid and hovers on spinning rotors. Semantic and motion-based discrimination attacks the false-alarm problem at its source instead of papering over it with sensor fusion after the fact.
It's software-defined and scalable. Cameras are commodity hardware. The intelligence lives in the model, which means capability improves through software updates rather than forklift hardware upgrades. New threat type? Retrain and redeploy. That is a profoundly different, and cheaper, cost curve than the radar-and-effector hardware stack, and it's the only economic model that can answer a thousand-dollar drone without spending a fortune per intercept.
So why hasn't vision-only already won? Because for most of C-UAS history, the camera couldn't be the brain. Wide-area visual scanning was computationally infeasible, classification wasn't reliable enough to trust without a human, and frame-based cameras were effectively blind between frames. The EO sensor needed a radar to point it and a person to interpret it.
That constraint is exactly what state-of-the-art AI removes. Modern object detection runs in real time. Edge inference puts that capability on a small companion computer at the sensor instead of in a distant data center. Pruning and optimization can shrink detection models by roughly a fifth or more while improving accuracy, so they fit and run fast on embedded hardware; recent pruned and quantized detectors report end-to-end latency around 36 ms on a Jetson-class edge device, well within the tens-of-milliseconds budget that small, fast targets demand. Emerging neuromorphic, event-based vision sensors largely eliminate the blind-spot-between-frames problem by reporting changes with microsecond temporal resolution, and fine-grained motion analysis can read the oscillation signature of a target: small drones' propellers spin at roughly 1,000 to 6,000 rpm, producing frequencies that are uncommon for birds and other airborne objects, which lets a system separate machine from animal.
In other words, the camera is finally allowed to graduate from witness to detective. The bottleneck was never the lens. It was the brain behind it, and that brain now exists.
Part 5: What a state-of-the-art aerial-threat brain actually has to do
Making the case for vision-only detection is the easy part. Building the brain that delivers on it is the hard part, and it's worth being concrete about what "state of the art" means here. A serious vision-first C-UAS perception system has to chain several capabilities together, in real time, on the edge:
- Detect small, fast, low-contrast objects against cluttered and changing skies, the genuinely hard end of object detection, where targets can be a handful of pixels.
- Track each candidate persistently across frames, holding identity through occlusion and manoeuvre, typically with a motion-estimation filter such as an extended Kalman filter that predicts where the target goes next.
- Classify what it's looking at with enough confidence to act: drone versus bird versus crewed aircraft versus authorized platform, and to do so with a false-alarm rate low enough that operators keep trusting the alerts.
- Decide and escalate within a latency budget tight enough to matter against a target crossing the sky in seconds.
- Integrate into the operational picture, emitting standardized situational-awareness output (for example, Cursor-on-Target / TAK messages) so the detection plugs into existing command-and-control, video-management, and response systems rather than living in a silo.
The unglamorous metric that decides whether such a system is actually useful is the false-alarm rate, especially the bird false-alarm rate. A detector that fires on every gull is a liability; a classifier that can confidently clear a bird is what turns raw detection into trustworthy intelligence. This is the benchmark on which vision-first systems must be judged, and the one that, handled well, becomes their defining advantage.
The bottom line
The drone threat is escalating, the legal authority to respond is finally arriving, and the capital is pouring in. Against that backdrop, the detection layer is quietly becoming the decisive battleground, and the threats are evolving precisely toward the kind of silent, autonomous, spectrum-decoupled machines that defeat the radio-frequency and even radar approaches the industry has leaned on.
What cannot be switched off is light and shape. A drone reflects photons and moves a body through space, and those are the signals a vision system reads. For most of the field's history, the camera lacked the intelligence to act on them alone. State-of-the-art AI changes that, turning a passive, low-cost, software-defined sensor into an autonomous brain that can see an aerial threat, name it, track it, and hand it off for response.
That is the case for an AI brain that detects aerial threats with vision alone. Not as a confirmation layer bolted onto radar, but as the primary sense for a generation of threats designed to be invisible to everything else