Client
Confidential Client · UAV Technology Program
Service
Product & Project Delivery, Technology Leadership, Strategic Advisory
location
International program context
Some engagements can be named. This one cannot. What can be described is the work: the direction of a drone technology programme built on two demanding engineering layers, video-based image recognition and the tracking and control systems that turn recognition into flight behaviour. The client, the deployment context and the commercial terms remain confidential. The technical and delivery story does not need them to be useful to a buyer.
What made this programme possible was not a contractor network assembled for the project. It was the Szel & Partners in-house engineering team, with fifteen years of continuous drone development behind it. People who have been building, flying, failing and rebuilding UAV systems long before AI-driven autonomous flight became a procurement category. That team is what allows the practice to take drone programmes from computer vision research through tracking and control technology to a working system, rather than stopping at a demonstrator or handing the difficult layer to a supplier.
The Challenge
Drone programmes fail in a predictable place. Image recognition that performs well on curated datasets degrades the moment it meets real flight: changing light, motion blur, vibration, scale shifts as altitude changes, and targets that leave and re-enter the frame. A detection model is not a capability. The hard part is the distance between a recognition result and a control decision the airframe can act on, in real time, with the latency and reliability budget that real operation allows.
The second challenge was organisational rather than technical. A programme of this type sits across three disciplines: computer vision, tracking and control engineering, and airframe integration. Each with its own development rhythm and its own definition of 'done'. Without a single technical direction holding the architecture, the interfaces between those disciplines become where the schedule and the budget disappear. The engagement was scoped to own exactly that: the architecture, the sequencing, and the delivery direction across all three layers.
The Approach
Phase 1: Architecture and technical direction. The programme started with the decisions that are expensive to reverse: how the vision layer, the tracking logic and the flight control stack would talk to each other, what ran on board versus off board, and what the real-time performance envelope had to be for the system to be operationally useful rather than merely demonstrable. Those boundaries were set in writing before development sprints began, which is what later allowed the three disciplines to build in parallel without integration collapsing at the end.
Phase 2: Video-based image recognition under real flight conditions (with AI technology). The computer vision layer was developed and hardened against the conditions that actually break detection in the air: variable illumination, motion and vibration, changing target scale, occlusion and re-acquisition. The measure of progress was never benchmark accuracy on a static dataset. It was whether recognition held up long enough, and stably enough, for the tracking layer downstream to rely on it.
Phase 3: Tracking and control technology. The final and most demanding layer turned recognition output into continuous tracking behaviour and control commands: maintaining a target across frames, handling loss and re-acquisition, and translating all of it into flight control input within a hard real-time budget. This is the layer where most drone vision projects stop, and it is the layer this programme was built to reach. Development stayed inside the Szel & Partners engineering team throughout, which is what let the capability compound into the practice rather than dispersing into a supplier chain.
The Outcome
The programme delivered what drone technology work is usually measured by: a video recognition layer that performs under real flight conditions, and a tracking and control capability built on top of it rather than promised alongside it. The technical architecture, the development direction and the engineering capability were held throughout by the Szel & Partners in-house team, which is what makes this kind of delivery repeatable across further programmes rather than being a one-off achievement.
What this case demonstrates is the part of Szel & Partners that does not show up in a strategy deck. Behind the advisory practice sits an in-house engineering team with fifteen years of continuous drone development, capable of taking an AI-driven programme from computer vision research through autonomous tracking and control technology to a working system. For founder-CEOs and corporate buyers whose problem is genuinely technical, that is the difference between receiving a recommendation and receiving a capability.
"In drone programmes the detection model is the easy half. The capability begins where recognition has to become a control decision, in real time, on a moving airframe. That is the layer we built, and the team we have been building for fifteen years is the reason we could."
—
Greg Szél
,
Founding Principal, Szel & Partners
Case Studies
