A Deep Learning Model for Drone Detection with Visualization

  • Zayed AlMessabi

Student thesis: Master's Thesis

Abstract

Recently, drone emerging issues have been all over the world. For example, In the past three months there were several incidents in different airports in the world. Drastic countermeasures have to be in place to reduce or eliminate associated risks. Regulations were established to control and regulate the increase possession of drones by normal civilians, but the affect was minimal. If there is no radical solution for this phenomenon it will continue in growing into an uncontrollable problem. This research thesis shows how modern visual detection technologies can be used to help recognize drones in the airspace. Since small drones cannot be detected with normal radars, visual solutions were mandatory to be in place. This research will present a new method for visual detection using a branch of machine learning called deep learning using convolutional neural network (CNN). This thesis presents a deep learning model and compares it with other available solutions. It will present in details how to tackle the problem mentioned earlier and how to fine-tune deep learning models for better detection accuracy. It will also present a visualizations technique for the deep learning models to show how these models work using filter visualization and heat maps with Gradient-weighted Class Activation Mapping (Grad-CAM).
Date of AwardMar 2019
Original languageAmerican English

Keywords

  • Drones
  • Breaches
  • Deep Learning
  • Machine Learning
  • CNN
  • Grad-CAM.

Cite this

'