Vision-Based Amateur Drone Detection: Performance Analysis of New Approaches in Deep Learning

Interest in unmanned aerial vehicles (UAVs) has increased significantly. UAVs capable of autonomous operations have expanded their application areas as they can be easily deployed in various fields. The expansion of UAVs' areas of operation also brings safety issues. Although legally prohibited places for UAV flights are defined, measures should be taken to detect violations. This study tested recently proposed methods that are used to detect objects from images on UV images, and their performances were discussed. We tested the models on a new dataset named GDrone that we created by collecting various images of drones. Two tested models, YOLOv6 and YOLOv7, have never been tested with a drone dataset. According to the experimental tests, the most successful model was YOLOv7 architecture, and its mAP (mean Average Precision) was 95.8% on GDrone dataset.