Real-time Adaptation of Drone Altitude and Object Detection Model for Moving Target Tracking

Persistent coverage control lets a drone patrol a field and keep watch on a target it finds there. Earlier work handled this with control barrier functions — one constraint keeping coverage of the field above a performance level, another keeping a detected target inside the drone’s sensing range. It works, but only because the target is assumed to stand still.

Move the target and the assumption breaks.

Illustration of the target tracking problem: drones at fixed altitude covering a ground plane with a moving target — Drones patrol the field under persistent coverage control; when one detects the target, it switches from patrolling to keeping the target in view.

Three things that decide whether tracking survives

Target speed is the obvious one — computation and motion delays eventually lose a fast target. The other two are the interesting part, because they pull against each other:

Detection model. Six TensorFlow detectors were tested, from MobileNet V1 (30 ms, 21 mAP) to Faster R-CNN Inception ResNet V2 (620 ms, 37 mAP). Accuracy and latency trade directly, and the most accurate model is not the best tracker — its latency loses the target anyway.

Altitude. Higher widens the field of view, so the target is less likely to slip out of frame. Higher also shrinks the target in pixels, which hurts detection. The right altitude therefore depends on which detector is running, and vice versa.

Tracking performance against drone altitude, plotted per detection model, at four target velocities — 96 experimental runs — six models × four altitudes × four target speeds. Below 0.1 m/s almost everything works. At 0.2 m/s the lightweight models start failing above 1.2 m, and the heaviest model underperforms at every altitude purely on latency.

Adaptation

Since those 96 measurements map each (altitude, model, target speed) triple to a tracking score, the choice at runtime becomes a small discrete optimisation over 4 × 6 = 24 options: estimate the target’s velocity, then pick the lowest altitude whose configuration still clears a performance threshold. Lowest, because a lower altitude gives more pixels on target — useful for whatever recognition or decision comes after tracking.

Experiment

Flown on the Tokyo Tech Robot Zoo Sky testbed against a Scamper O-308 ground robot carrying an A3-size printed car, accelerating from 0.1 to 0.3 m/s over about two minutes.

Tracking accuracy over time without and with the adaptation algorithm, alongside target velocity, detection model and altitude — Left, fixed configuration: the drone loses the target at about 65 s, as soon as it reaches 0.2 m/s. Right, with adaptation: at the same moment the drone climbs to 1.2 m and switches to a heavier detector, and accuracy stays near 1.

Snapshots of the experiment without the adaptation algorithm, drone camera view inset — Without adaptation. The camera inset holds a clean detection through 50 s; by 65 s the target is sliding out of frame, and at 80 s the inset shows bare floor.

Snapshots of the experiment with the adaptation algorithm, showing the drone camera view and the coverage state — With adaptation. The coloured field is the coverage state; the drone stays with the target as it speeds up.

The altitude set is discretised here, and target speed is sampled at four values. Extending both to continuous ranges is the stated next step.

Carried out with support from Fujitsu Limited.