F2 Інженерія програмного забезпечення

Permanent URI for this collection

https://ekmair.ukma.edu.ua/handle/123456789/18529
Освітня програма: "Інженерія програмного забезпечення"

Browse

Browsing F2 Інженерія програмного забезпечення by Author "Kurochkin, Andrew"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Development and Implementation of a Military Technology Trends Monitoring System
    (2025) Prokhorov, Oleksandr; Kurochkin, Andrew
    This work presents the design and implementation of a system for monitoring technological trends in the military sector using Telegram as a data source. The system automatically collects, processes, and analyzes both historical and real-time posts from selected Telegram channels, focusing on the emergence and dissemination of key terminology such as "реб" (eng.: "electronic warfare") in our evaluation case study. A modular architecture was developed, combining Go-based data scraping, Python-based aggregation and keyword analysis, and a Grafana dashboard for visualization. The system supports both local Docker-based deployment and cloud-based deployment via Terraform on AWS. Evaluation included performance benchmarks, peak resident-set size (RSS) profiling, and a case study comparing our system’s findings against professional media and Google Trends. Results indicate that a Telegram-based pipeline can detect rising interest in electronic-warfare topics earlier than traditional information channels.
  • Thumbnail Image
    Item
    Training YOLO Models for Real-Time Object Detection on UAV
    (2025) Solovei, Tymofii; Kurochkin, Andrew
    In this study, we have explored the implementation of the YOLOv8(nano) model for the task of real-time detection of military objects for UAV companion computers. We have collected and merged different datasets from the open sources with clearly annotated classes such as tanks, armored vehicles, armored personnel carriers, etc. Additionally, datasets with civilian people and vehicles have been included to address ethical concerns. Because of the poor quality of original datasets, we developed a processing pipeline for proper data selecting, filtering, and augmentation. We trained the YOLOv8-nano model for 100 epochs. The default pre-trained on the COCO dataset YOLOv8-nano model initially achieved an mAP@0.5 of 0.305 and mAP@0.5-0.95 of 0.169 on our dataset. Our final YOLOv8-nano model achieves a mAP@0.5 of 73.61% and mAP@0.5-0.95 of 51.18%. We also evaluated our model using combat videos from FPV(First Person View) drones containing different military targets.