Deep Learning Based Malicious Drone Detection Using Acoustic and Image Data

EasyChair Preprint 9335

Abstract

Autonomous drones have been studied in a variety of industries including delivery services and disaster protection. As the supply of low-cost drones has been increasing, a CUAS (Counter unmanned aerial systems) is critical to manage autonomous drone traffic control and prevent drone flights in secured areas. For these systems, drone detection is one of the most important steps in the overall process. The goal of this paper is to detect a drone using the microphone and the camera by training deep learning models based on image and acoustic features. For evaluations, three methods are used: visual-based, audio-based, and the decision fusion of both features. The decision fusion of audio and vision-based features is used to obtain higher performance on drone-to-drone detection. Image and audio data were collected from the detecting drone, by flying two drones in the sky at a fixed Euclidean distance of 20m. In addition, deep learning methods are applied to investigate an optimal performance. CNN (Convolutional Neural Network) was used for acoustic data, and YOLOv5 was used for computer vision. From the result, the decision fusion of audio and vision-based features showed the highest accuracy among the three evaluation methods.

Keyphrases: Collision Avoidance System, UAV detection, audio classification, computer vision, deep learning

@booklet{EasyChair:9335,
  author    = {Juann Kim and Dongwhan Lee and Youngseo Kim and Heeyeon Shin and Yeeun Heo and Yaqin Wang and Eric T. Matson},
  title     = {Deep Learning Based Malicious Drone Detection Using Acoustic and Image Data},
  howpublished = {EasyChair Preprint 9335},
  year      = {EasyChair, 2022}}