Optimization of YOLOv8 architecture for UAV object capture tasks: analysis of the trade-off between accuracy, speed and computational resources
- Authors: Satsuk A.V.1
-
Affiliations:
- Donetsk Institute of Railway Transport (DonIRT), Chair "Automation, Telemechanics, Communication and Computer Engineering"
- Issue: No 2 (2025)
- Pages: 35-42
- Section: Information technology, automation and telecommunications
- URL: https://journals.rcsi.science/0201-727X/article/view/363095
- DOI: https://doi.org/10.46973/0201-727X_2025_2_35
- ID: 363095
Cite item
Full Text
Abstract
This paper presents a comprehensive approach to optimizing the YOLOv8n model for object detection tasks using unmanned aerial vehicles (UAVs) under constrained computational resources. The focus is on quantization (INT8/INT4) and pruning (50%/75%) techniques aimed at reducing the model's computational complexity while maintaining acceptable accuracy. As a result of optimization, the YOLOv8n-Optimized-Drone model was developed, demonstrating a 4-fold increase in processing speed on the Raspberry Pi 5 platform compared to the basic version. The model size was reduced by 3.8 times, which is critical for embedded UAV systems.
A specialized dataset with bounding box markup was created for training and validating the model, taking into account the UAV shooting conditions. Field tests confirmed the effectiveness of the proposed method, which provides a balance between performance, power consumption, and accuracy. Additionally, the influence of different quantization and pruning levels on final metrics was investigated, enabling the determination of the optimal configuration for deployment on low-power devices. The obtained results open prospects for further adaptation of the model to dynamic flight conditions and integration with multi-sensor UAV systems.
About the authors
A. V. Satsuk
Donetsk Institute of Railway Transport (DonIRT), Chair "Automation, Telemechanics, Communication and Computer Engineering"
Author for correspondence.
Email: alexandrsatsuk@gmail.com
Candidate of Engineering Sciences, Associate Professor
References
- Улучшенный алгоритм на основе YOLOv5 для обнаружения амброзии полыннолистной на изображениях, полученных с помощью БПЛА / С. Сяомин, С. Тяньцзэн, М. Хаомин [и др.] // Front. Plant Sci. – 2021. – Т. 15. – С. 1360419. – doi: 10.3389/fpls.2024.1360419.
- Real-time Object Detection for UAV Aerial Images Based on Improved YOLOv7 / Y. Zhang, X. Chen, H. Li, Y. Wang // Journal of Electronic Imaging. – 2015. – Vol. 32, No. 6. – P. 063002. – doi: 10.3390/electronics12234886.
- Сацюк А. В., Белый Р. В., Ищенко А. Е. Оценка эффективности алгоритмов YOLO для обнаружения объектов в реальном времени во встраиваемых системах беспилотных транспортных средств // Сборник научных трудов Донецкого института железнодорожного транспорта. – 2024. – № 4 (75). – С. 73–82. – ISSN 1993-5579.
- Redmon J. A., Farhadi A. YOLO9000: Better, Faster, Stronger // Conference on Computer Vision and Pattern Recognition (CVPR). – 2017. – arXiv:1612.08242v1.
- Сацюк А. В., Воевода Е. Г. Анализ трекеров алгоритмов компьютерного зрения в вопросах отслеживания подвижных объектов в видеопотоке // Сборник научных трудов Донецкого института железнодорожного транспорта. – 2023. – № 71. – С. 54–65. – ISSN 1993-5579.
- Библиотека программиста. JavaScript для глубокого обучения / Ф. Шолле, Э. Нильсон, С. Бэйлесчи [и др.]. – Санкт-Петербург: Питер, 2021. – 576 с. – ISBN 978-5-4461-1697-3.
- YOLOv8. Ultralytics. Официальный репозиторий YOLOv8 / G. Jocher, A. Stoken, A. Chaly [и др.]. – 2023. – URL: https://github.com/ultralytics/ultralytics
- (дата обращения: 07.02.2025).
- Zheng Z., Wang P., Ren D. [et al.]. Enhancing Geometric Factors in Model Learning and Inference for Object Detection and Instance Segmentation // Transactions on Cybernetics. – doi: 10.1109/TCYB.2021.3095305.
- An Efficient UAV-Based Aerial Image Object Detection Method via Improved YOLOv4 / B. Liu, X. Li, G. Wang [et al.] // Remote Sensing. – 2023. – Vol. 12, No. 21. – P. 3599. – doi: 10.1109/SIU53274.2021.9478027.
- Сацюк А. В., Воробьев А. А. Особенности разработки беспилотных летательных аппаратов для отрасли железнодорожного транспорта // Сборник научных трудов Донецкого института железнодорожного транспорта. – 2023. – № 68. – С. 13–21. – ISSN 1993-5579.
- Сацюк А. В., Швалов Д. В. Автономное наведение БПЛА с использованием компьютерного зрения: проблема точного управления рулями // Автоматика на транспорте. – 2024. – Т. 10, № 4. – С. 372–381. – doi: 10.20295/2412-9186-2024-10-04-372-381.
- Lalak M., Wierzbicki D. Automated Detection of Atypical Aviation Obstacles from UAV Images Using a YOLO Algorithm // Sensors (Basel). – 2022. – Vol. 22, No. 17. – P. 6611. – doi: 10.3390/s22176611.
Supplementary files
