Date Approved
7-27-2023
Embargo Period
7-28-2023
Document Type
Thesis
Degree Name
Master of Science
Department
Electrical and Computer Engineering
College
Henry M. Rowan College of Engineering
Advisor
Nidhal C. Bouaynaya, Ph.D.
Committee Member 1
Gregory Ditzler, Ph.D.
Committee Member 2
Thomas Kiel
Keywords
Vision, Drone, Object Detection, Virtual Reality
Subject(s)
Machine learning; Military engineering
Disciplines
Electrical and Computer Engineering | Engineering
Abstract
The introduction of aerial drones on the modern battlefield has transformed combat operations, posing a significant threat to ground-based military operations. Detecting drones in safety scenarios is crucial. However, modern machine learning (ML)-based object detectors struggle to detect small objects like drones. This thesis presents three main contributions: (a) data and algorithmic modifications to improve small object detection in YOLO to aid in drone detection, (b) the development of a benchmark drone detection dataset called DyViR, and (c) the implementation of explainable artificial intelligence (XAI) to ensure transparent and trustworthy decision-making. To boost the performance of small object detection, we introduce the Normalized Wasserstein distance (NWD) into the loss function of our ML model. By incorporating this distance metric, we can effectively handle small object detection by reducing the penalty assigned to small objects, and appropriately balancing the significance of different object sizes. This allows the model to prioritize the accurate detection of small objects, enhancing overall performance. To evaluate our algorithm, we developed and tested the DyViR dataset specifically designed for drone detection research. This synthetic dataset provides a benchmark for assessing drone detection performance. In combat settings, the trustworthiness of ML systems is paramount as their decisions impact user survival. Thus, we implemented explainable AI systems (XAI), specifically Grad-CAM and Eigen-CAM, These techniques provide explanations for the models’ decisions, increasing trust in the system for developers and users.
Recommended Citation
Koutsoubis, Nikolas, "MACHINE LEARNING-BASED DRONE AND AERIAL THREAT DETECTION FOR INCREASED TURRET GUNNER SURVIVABILITY" (2023). Theses and Dissertations. 3146.
https://rdw.rowan.edu/etd/3146
