K-PARD · An Alterra Solutions R&D project

Keeping a fast drone in sight

K-PARD detects small, fast drones in camera images, tracks them and follows them with a moving turret. All processing runs on a computer in the field; no internet connection or cloud service is needed.

The tracking software screen: a drone flying against a cloudy sky, with a detection box and confidence score
A frame from the first field trial (May 2026). The software finds the drone in the sky, boxes it and follows it. Taken from a screen recording of the tracking software; some on-screen labels were removed.

Why is it hard?

Most drones in the published drone-detection datasets hover or drift past slowly. A fast FPV drone covers dozens of times its own size every second. Common methods can't hold on to a target at that speed.

37×

In the FPV profile we design for, a drone moves about 37 times faster, relative to its size, than the typical drone in public datasets.

Boxes stop overlapping

Common trackers recognise a target by the overlap of its boxes in two consecutive frames. At this speed the boxes no longer overlap and the track breaks.

Zooming doesn't help

Zoom or higher resolution makes the drone bigger, but its jump from frame to frame grows by the same factor. The problem is not the camera; it is how the target moves.

How it works

01

Motion detection

Picks up small moving points in the sky as candidates.

02

AI verification

An object detection model looks only at the area around each candidate and checks whether it really is a drone. False alarms are filtered out here.

03

Tracking

The target is followed from frame to frame by tracking image motion and predicting its next position. If it disappears for a few frames, the prediction keeps following it.

04

Turret

A two-axis turret keeps the target in the centre of the frame. If the link to the computer drops, it holds its position and does not move on its own.

Closed network. Needs no internet and no cloud service.
Swappable parts. Detection methods plug in and out; a new sensor joins the same pipeline.
Ordinary camera. Today's test rig runs on a phone camera.

Measure first

Every camera frame is recorded together with the angle sent to the turret at that moment, so it is always possible to see afterwards what the system was following and when.

Tracking methods are tested on synthetic footage with controlled, increasing speed and on published real-world datasets, and we measure at what speed each one breaks. A change goes into the system only if it improves that measurement.

Where it stands

K-PARD is an R&D prototype. The measurement pipeline and the live tracking loop work, and the first field trials are done. Development continues towards the speed and accuracy we are aiming for.

If you work in defence or security and are considering a field trial or joint development, get in touch.

Contact us