Date of Award

3-2024

Document Type

Thesis

Degree Name

Master of Science in Computer Science

Department

Department of Electrical and Computer Engineering

First Advisor

Richard Dill, PhD

Abstract

sUAS present significant risks to local and federal agencies when under the control of negligent, reckless, or criminal operators. In the face of an escalating presence of sUAS in shared airspace with traditional aircraft, and their deployment in protected airspace as potential weapons, safeguarding personnel, facilities, and assets becomes paramount. This research seeks to address this emerging threat by investigating the efficacy of integrating low-cost distributed sensors and Machine learning (ML) models to enhance battlespace awareness and complement existing sensing platforms for real-time sUAS detection, classification, and localization. The thesis introduces the conceptualization and development of a Drone Detection Command Center (DDCC). The DDCC interfaces with a distributed node system de- signed to acquire sUAS data in real-time through both visual and audio modalities. Furthermore, the DDCC possesses the ability to capture frames of interest, facilitating their integration into future machine learning models for enhanced prediction capabilities. Data is collected of the DJI Matrice 600 Pro using the system and the data is used to create multiple deep learning models with the goal of classifying sUAS presence and predicting the range the sUAS is from a given node. A focus is placed on evaluating the performance of range predictions based on audio, then comparing those to the range predictions based on video. Finally, the data is fused into a single dataset, and the same predictions are made on a custom model with the goal of determining if fused data will present superior results to individual modalities. As an initial step, audio classification achieves a categorical accuracy of 79.6%, while video classification achieves an accuracy of 86.7%. From there, range predictions are iv made with audio and video datasets independently, providing a mean absolute error of 10.463 meters for audio, and 16.961 meters for video respectively. Finally, the audio and video data is fused together and fed into a Convolutional Recurrent Neural network to test if the combined data will provide better results. The combined results show that the mean absolute error is 9.57 meters, an improvement of .88 meters over audio data, and 7.385 meters over video data.

AFIT Designator

AFIT-ENG-MS-24-M-182

Comments

A 12-month embargo was observed for posting this work on AFIT Scholar.

Distribution Statement A, Approved for Public Release. PA case number on file.

Download

Share

COinS