The next phase of artificial intelligence (AI) at the edge for defense applications will not be defined simply by whether a model can run onboard. It will be defined by whether detection, classification, and tracking can work together quickly and reliably enough to preserve understanding as the mission unfolds. For first-person-view (FPV) and counter-uncrewed aerial systems (UASs), this goal will mean maintaining continuity on relevant objects, reducing unnecessary operator workload, and supporting faster decisions on platforms constrained by size, weight, power, and cost (SWaP-C). It also creates a path toward more supervisory forms of control, in which operators can oversee more systems without surrendering responsibility for consequential decisions. The operational advantage is therefore not AI at the edge by itself; rather, it is the ability to give smaller platforms enough onboard understanding to identify what matters, sustain attention on it, and present useful information when the operator needs it most.
As global defense and security organizations rapidly adopt counter-uncrewed aerial system (UAS) strategies and technology, they face a persistent problem. An attacking drone only needs to get through once to be successful, while an interceptor must repeatedly detect, track, pursue, and engage multiple fast-moving targets.
That challenge was brought into stark relief when NATO personnel and European defense companies recently tested drone-interceptor systems at Latvia’s Sēlija training area in May of 2026: The demonstration included successful intercepts and misses, highlighting how difficult it remains to reliably neutralize small, agile aerial threats.
Counter-UAS illustrates a broader technical challenge across intelligence, surveillance, and reconnaissance (ISR); loitering munitions; and other defense applications for drones. Detecting an object is just the start: A system still needs to determine what the object is, whether it’s important, whether it’s the same object identified moments earlier, and whether it can track the object long enough to support an effective response.
Whether to run artificial intelligence (AI) onboard is not the primary technical question. The more pressing part is turning detection into reliable understanding that is relevant to the mission as it unfolds in unpredictable ways.
Onboard detection is just the starting point
The operational value of edge AI depends on what happens after the initial detection, ideally occurring at the maximum possible range to cue countermeasures. Depending on the platform and mission, the system may need to detect, classify, track, and prioritize objects in real time while generating information for an operator or another mission system.
Classical computer-vision techniques are required for initial detection as they require far fewer pixels for initial assessments and raising the initial alarm of a possible threat.
From there, modern AI-assisted target recognition rarely relies on a single model making a single decision. Increasingly, developers are deploying multistage AI pipelines in which one model identifies potential objects, additional models classify them (such as birds, civilian aircraft, crewed aviation, etc.), and tracking algorithms maintain continuity across successive frames. Breaking the process into stages increases confidence and reduces false positives, enabling operators to focus on fewer, higher-confidence events.
The precise sequence will vary depending on the mission, sensors, platform, and available compute, but the goal will always be to advance from the initial detection to information that remains useful even as the scenario changes in real time. A mission-ready system (especially for defense and security) must continue producing actionable information as the platform moves, the object(s) change orientation, and the background (sky, ground, water) becomes cluttered with images. A system that performs well on individual image frames but loses continuity as things change may lose track, resulting in consequential outcomes.
Edge AI performance should therefore be evaluated as a workflow and not as a collection of siloed model scores. Detection accuracy matters, but so do factors like classification confidence, tracking persistence, latency, false-positive rates, and how information is presented to the operator. These elements determine whether AI reduces cognitive load for the operator or adds another layer of data to interpret when milliseconds matter (such as for seekers).
For developers and integrators, that part means evaluating more than detection accuracy: They should also measure how long the system maintains a track, how quickly it recovers after an object is obscured, how confidence changes as conditions deteriorate, and whether the full processing chain performs consistently under realistic operating conditions. And don’t forget the rookie rule: If confidence is 1.00, it is a test video trained primarily on viewed data.
The architecture should begin with the use case. A counter-UAS system monitoring a massive area of sky might need to be trained with many different objects at different horizon levels, prioritizing maximum range detection. In contrast, an FPV drone following an object may place greater emphasis on persistent low-latency tracking and rapid response to changing perspective and orientation. In both instances, the models and processing stages should be selected to support the mission.
Tracking turns detection into persistent understanding
Identifying an object correctly once is not enough. Drone operations unfold rapidly across continuous video, not in isolated frames. The camera and the object it is tracking may be moving simultaneously as distance, perspective, lighting, and terrain change in real time.
Detection answers whether an object is present, while classification helps determine what the object is. Tracking answers a fundamentally different question: Is this still the same object, and can the system continue to follow it as conditions change?
Today’s AI can support aided target recognition (AiTR), maintain persistent tracks across successive frames, classify objects with increasing confidence, and prioritize detections based on mission requirements. These capabilities help preserve continuity after the initial detection rather than forcing the system or operator to interpret every frame as a new event. The best trackers even use AI to improve track quality and help recover lost tracks.
Such continuity by adding AI matters when an object moves behind terrain, enters visual clutter, changes direction, or becomes challenging to distinguish from similar objects. If that relationship is lost, the system may generate duplicate alerts. In practice, the operator may then have to reacquire the object manually, might get overloaded with their battle-management system, or simply be unable to guide to a threat.
Pixel lock refers to maintaining track on a pixel/s after an object is identified, while tracking may also enable detection, classification, and recognition to operate as a coordinated workflow rather than treating each video frame as a new problem. These processes are especially critical on small drones with limited compute power and thermal capacity that cannot run AI on every frame while still achieving low latency for guidance and control.
The value of tracking is therefore not just about keeping an object on screen. Rather, it’s about preserving sufficient continuity to reduce repeated manual reacquisition, support faster interpretation/guidance, and create a foundation for human oversight.
Designing for low-SWaP platforms
The shift from detection to persistent understanding matters only if the complete processing chain can operate within the limits of the platform. On small UASs and interceptors, recognition accuracy should be balanced against frame rate, latency, size, weight, power, thermal limits, and cost.
Just a few years ago, sophisticated computer vision often required dedicated GPU hardware that limited deployment to larger aircraft. Updated lightweight AI accelerators now enable many of those same capabilities to run on compact drones and other SWaP-constrained platforms that could not support them previously and cost less than the previous versions.
The engineering goal is not to run the largest possible model but rather to choose the right combination of detection, classification, and tracking capabilities for the mission. The next step is to sequence them in a way that preserves performance without overwhelming the available compute power.
The design tradeoff is not just about accuracy versus speed. Engineers must choose which functions must run continuously and how much processing can be sustained without exceeding the platform’s power, memory, and thermal limits. The result is not simply faster AI; it is more practical AI that can operate across a wider range of defense platforms and mission profiles.
What edge AI changes for FPV and counter-UAS
For FPV drone operations, the value of edge AI lies in reducing the continuous manual-control burden required from the operator. Using edge AI, the operator doesn’t have to remain fixed on one video feed and correct every movement but can instead focus on confirming an object’s relevance, assess context, and intervene with confidence decreases or the mission changes. It also enables mission completion in contested communications environments where remote communications are often lost.
The shift creates a path toward higher operator-to-platform ratios, where one operator may supervise multiple systems instead of staying tied to a single drone. The goal is not to replace human judgment, especially when FPVs are used in a sensor-to-shooter configuration with an ISR drone. The goal is to ensure that humans focus their attention on decisions rather than data management and manual stick control.
For counter-UAS, there’s a slightly different challenge. Detection may show that something is there in the sky, but classification needs to distinguish a drone from a bird and a friendly asset that does not require a response. The system needs to preserve enough continuity to support pursuit and interception.
This aspect is especially important for lower-cost interceptors. As AiTR capabilities become more available on smaller, lighter processors, systems that used to require heavier, more expensive computing may become practical on drones designed for a more sustainable cost exchange. The goal is to provide the interceptor with enough onboard understanding to quickly track and pursue the right object.
Stephen Bornstein is Managing Director of Australia and SVP of Product and Engineering for Sightline Intelligence. He leads the company’s global product strategy for AI-powered computer vision and autonomy solutions supporting defense and national security missions. A founder of Athena AI, Stephen is an award-winning aerospace engineer with deep expertise in defense technology, robotics, AI, and product innovation.
Sightline·https://sightlineintelligence.com/
Featured Companies
