Deep Learning Inspired Vision based Frameworks for Drone Detection

Abstract

Drone detection technology is a new frontier in defence systems. With increasing incidences of crimes and terroristic attacks using commercial drones, detection of unauthorized drones is critical for timely responses from law enforcement agencies. In this paper, the issues of unavailability of benchmark dataset and performance metrics for drone detection are addressed and three single shot detectors, based on YOLOv4, YOLOv5 and DETR architectures are presented. A maximum of 99% average precision (AP) with an average Intersection over Union (IOU) of 84% was achieved. The precision-recall curves corroborate the generalization and fitness of the trained detection models.