Person Detection in Drone Imagery

The use of drones in search and rescue operations has become standard almost everywhere in the world. A special challenge in the search and rescue operation is the automatic detection of persons in different terrains, in different situations and body positions, in different weather conditions, and from different shooting heights during a drone flight. This paper investigates the accuracy of people detection in drone images on existing VisDrone, Okutama - Action datasets, and on a custom SARD image dataset built to simulate search and rescue scenes. A Faster R-CNN with FPN as the backbone, pre-trained on the COCO data set, was used as a detector. The person detector is additionally trained on the SARD data set containing 1,981 images and on the subset of the VisDrone set. After transfer learning, a significant improvement in the detection results of persons in the images taken by the drone was achieved concerning mAP and precision and recall.