Drone Detection using YOLOv4 and Amazon Rekognition

Abstract

Creating effective tools and techniques for the automatic detection of drones is the need of the hour. Several detection technologies are under development, with a variety in complexity, range, and capabilities. There are some gaps in the present surveillance models, as the compact size and fast maneuvers result in drones being a difficult target to be detected when compared to other aircraft. Effective usage of video analytics does solve the problem, but powerful algorithms are needed to operate also under unfriendly conditions such as long-distance, low visibility, etc. This study aims to detect one or multiple drones in video sequences where drones appear at the same point when other distractor objects may also be present, whereas, algorithms should provide analysis only when the drone is present and not resulting alarms on other distractors, nor rambling by the rest of the frame. In particular, two approaches based on different deep learning strategies are put-forward and compared on a substantial dataset. Experimental analysis, as well as visual analysis of the proposed technologies, are discussed in the paper.