Developing Robust AI Models for Identification Friend and Foe

Abstract

Accurate and timely identification of friendly and foe (IFF) assets, particularly tanks, is critical in modern military operations to prevent friendly fire and support effective decision-making. Existing methods often suffer from inefficiencies, delayed response times, and difficulty adapting to changing battlefield conditions or distinguishing between visually similar tank models. This thesis addresses these issues by developing and evaluating an AI-based system for real-time tank detection and classification. The system uses a two-stage deep learning pipeline: a fine-tuned YOLOv8 model first detects tanks in video feeds, followed by classification of the detected regions using one of three convolutional neural networks (CNN) architectures either ResNet-50, MobileNetV2 or a custom-designed CNN to determine tank type and affiliation. Models were trained and tested on a custom dataset of four tank classes, with data augmentation to improve balance and visual diversity. Results show that the YOLOv8 detector achieved over 99.5% precision and recall, while classification accuracy reached % with ResNet-50, 99.94% with MobileNetV2, and97.39%with the custom CNN, each offering different inference speeds. The system was implemented as a standalone desktop and a web application, demonstrating real-time IFF capabilities. This work contributes a practical and adaptable framework for accurate tank identification, offering potential improvements in battlefield safety and operational awareness.

Date of Award	2025
Original language	American English
Supervisor	Chan Yeun (Supervisor)