Drone detection of armed people among civilians

The camera is on a drone flying over dry farmland. Dirt tracks cross the fields, scrub grows between them, and a village of flat-roofed houses sits at the horizon. The drone keeps moving during the clip, so the horizon and the fields shift in the frame. A detector therefore has to cope with a changing background and not only with moving people.

Three people walk along one of the tracks. Each is boxed in green and labelled civilian. Partway through, a group of five or six armed figures enters from the left. They are boxed in red and labelled armed military. The clip is built around one question that matters in this kind of footage: telling people who are armed from people who are not, at a distance and from above.

People seen from a drone are small in the frame, and a rifle is a thin object at that size. The labelled boxes show how large the targets are and how close the two groups are to each other. The green and red boxes are a trained detector's output drawn onto the clip for illustration. The labels sold with the dataset are YOLO and COCO files.

The original drone footage, without the armed group, is shown lower on this page. The armed figures were generated into the empty scene. The dataset for this scenario has 800 images with civilian, armed_person and vehicle classes, and there is no free sample, so it is licensed or arranged through contact.