Airborne Localisation of Small UAS using Visual Detection: A Field Experiment

The development of Uninhabited Aerial System (UAS) technology has grown immensely in the last decade with its uptake producing many novel applications. However, the speed of technology adoption also brings its share of challenges, with restricted airspace breaches being a particular safety concern. From a military perspective, detecting and countering armed UAS and adversarial surveillance is also of high importance. To assist with these problems, this paper proposes a method of localisation of Small Uninhabited Aerial Systems (sUASs) commonly known as `drones' from other airborne sUAS using vision-based detection. The Convolutional Neural Network-based Single Shot Detector architecture for real-time object detection known as You Only Look Once (YOLO), chosen for its speed and accuracy, was trained on a custom drone dataset and used to derive absolute coordinates of detected drones from the local position of an observing drone. These coordinates can then be used for further observation and action. Implementation on a physical aircraft for demonstration purposes shows that a target drone can be localised using off-the-shelf components to within an error of 2. 62m (±0.70m 95% CI) when attempting to maintain a standoff distance of 12. 5m. Analysis on the performance of the YOLO algorithm presents a performance trade-off, where a higher YOLO input network resolution results in a lower distance error and a longer YOLO inferencing time.