Multi-Sensor Drone Detection
I fuse radar, sound and thermal to find small drones in the airspace where each sensor alone goes blind.
Overview
This began as my final-year university project (COMP1682), scoped with a formal proposal and a Gantt chart, and grew into a working build. The target is the hardest airspace there is: a small drone in the last hundred metres, low and fast against ground clutter. My rig fuses radar, an acoustic array and a thermal camera into one tracking pipeline, and as of July 2026 it detects a contact and holds a live track on it. It is detection and tracking research only, and that boundary is deliberate and permanent.
Why one sensor is never enough
The design started from a hard look at how each sensor fails down low. Radar is the only all-weather option with direct velocity measurement, but near the ground the clutter takes over: returns from buildings and terrain can be thousands of times larger than a small drone's radar cross-section, which is typically under 0.01 square metres, and quoted detection ranges collapse from kilometres in open ground to a few hundred metres in cluttered environments. An acoustic array is passive, hears in every direction and works through foliage where radar has no line of sight, but its practical range is roughly 120 to 180 metres, and identification accuracy in the research I worked from falls from over 90 percent inside 120 metres to around 60 percent at the edge of that range, with urban noise cutting it further. Thermal is excellent close in, because motors and electronics glow against the background, but it is line of sight only and fog can cut its range by up to 80 percent.
The point is that these failures do not overlap. The literature I based the design on puts single-sensor detection at 75 to 90 percent, two-sensor fusion at 95 to 99 percent, and three-sensor fusion above 99 percent. Fusing all three is not a nice-to-have at this altitude, it is the only approach that works.
The pipeline
Each sensor runs its own front end and produces detections at its own rate, so the first job is getting every measurement onto a common time base before fusion can even begin. The fusion stage then associates detections into tracks and runs Kalman-family state estimation over them. I studied both the extended and unscented variants: the unscented filter's sigma-point approach suits small multirotors because it handles their non-linear motion without linearising it. A YOLO-family classifier on the thermal and optical feed answers the question radar cannot: drone, bird or clutter. Everything runs on edge compute with no cloud dependency.
Power and duty cycle
I sized the power budget before committing to the operating design. A small radar in idle listen draws in the region of 100 watts, while a full-power tracking burst sits around a kilowatt but lasts only tens of seconds. Worked through, one tracking event per hour costs roughly 106 watt-hours, so a full day of listening is battery territory rather than generator territory. That arithmetic is why the rig sits in a low duty-cycle idle-listen state and only commits full power when a contact appears.
Where it stands
I started the build in June 2026 and had live tracking working by July: detection, a moving track on the fusion display, and classification overlays on the thermal and optical feeds. Refinement is ongoing. The scope stays fixed: this system finds and follows drones, nothing more.
The hardest problem
Time synchronisation, inside a window that barely exists. The radar, the acoustic array and the thermal camera all update at different rates, and fusion is meaningless until every measurement sits on one shared time base: associate a radar return with an acoustic bearing from a fraction of a second earlier and the track swings all over the place. The window itself is brutal too. A fast, low drone covers 100 metres in under two and a half seconds, and terrain can mask it until the last moment, so the whole detect, classify and track chain has to complete in well under a second. Getting measurements timestamped, aligned onto a common clock and associated to the right track quickly enough to hold a stable live track on a real target was the core of the project. It works now, and I am still tightening it.
Start a project
Want something built like this?
Everything on this page was designed, built and hosted by one person. Tell me what you need and I reply with scope, cost and timeline, usually within two working days.