Researchers at Macquarie University have built a drone-to-drone handoff system that identified the intended target in 96% of frames during an outdoor trial with several similarly dressed people. The system, named Perspective Alignment and Tracking Handoff, or PATH, also completed 42 of 45 end-to-end handoffs. The results were posted September 11 in an unreviewed preprint.
Long-duration tracking exposes a practical limit: a drone can run low on battery while its camera is still following a person, animal or vehicle. Another drone can take over, but it sees the scene from a different position and scale. Matching appearance alone becomes unreliable when several candidates look alike, while ordinary satellite positioning may lack the relative precision needed to distinguish nearby targets.
PATH passes a location between the drones instead of exchanging video. The sender uses a camera that measures depth to convert the centre of its tracked target into a three-dimensional point. The receiver observes a small reference marker on the sender, calculates where the two drones are relative to each other and projects the transmitted point into its own camera view. It then searches that area and sends a detected box back. The sender compares the returned box with its original target and transfers tracking only after their overlap holds for seven seconds.
The team tested custom quadrotors and a second pair of DJI Tello drones. An indoor measurement used three target locations and seven receiver positions, producing 420 observations. The outdoor ambiguity trial placed six volunteers in similar dark clothing and caps, then compared PATH with ORB and XFeat, two methods that match visual features across images. The researchers classified 350 frames by whether the receiving drone found the same person specified by the sender. Separate end-to-end trials followed a small car, a micro-drone and a walking person.
PATH handled 336 of the 350 outdoor frames correctly, including cases in which the specified person was absent, and recorded seven false selections and seven missed targets. ORB recorded a 62.1% false-negative rate, while XFeat recorded a 15.4% false-positive rate. In the indoor measurements, the mean relative-position error between drones was 4.7 centimetres and the mean target-position error was 3 centimetres. The handoff calculations ran on a Jetson Nano processor, with communication below 16 kilobytes per second at 60 updates per second.
The trial covered two drones that were already close enough to see each other. PATH still depends on the reference marker remaining visible, and an obstruction can interrupt the transfer; the authors also note that its communication could be jammed. Larger teams, outdoor missions with longer separations and markerless relative positioning remain untested. Those experiments will show whether the measured handoff accuracy survives the visibility, motion and network failures of sustained field work.
Sources
- Heegyeong Kim and colleagues, arXiv, September 11, 2026. Abstract, authorship, submission date, principal results and licence.
- Full paper. System sequence, experimental setup, comparisons, hardware measurements and limitations.
- Lead image: Figure 2 from Kim and colleagues, reproduced from the paper under CC BY 4.0.