Author:
Yermolenko Ruslan,Klekots Denys,Gogota Olga
Abstract
This study aimed to train algorithms for detecting commercial unmanned aerial vehicles using machine learning techniques. Neural network architectures YOLOv8 and MobileNetV3 were used to detect unmanned aerial vehicles in images and videos. The models used were pre-trained on the ImageNet dataset and then refined on the SimUAV dataset containing images of four types of drones (Parrot A.R. Drone 2.0; DJI Inspire I; DJI Mavic 2 Pro; and DJI Phantom 4 Pro), different sizes and in eight different background locations. The study confirmed that the combination of the YOLOv8 and MobileNetV3 architectures has significant potential for detecting commercial unmanned aerial vehicles in various types of images. The trained models demonstrated high performance in the recognition and classification of unmanned aerial vehicles, achieving an average detection accuracy (at an IoU threshold of 50%) of 0.747 and 0.909 for the MobileNetV3_Small and MobileNetV3_Large models, respectively. This demonstrates the high efficiency and accuracy of the models in detecting objects on the test data. The results of the study also included the values of the binary cross-entropy metric, which were 0.308 and 0.216, respectively, indicating the high accuracy of the models in object classification and confirming the high efficiency and reliability of these models in working with objects on the test data. During the study, the MobileNetV3_Large model showed more accurate results than MobileNetV3_Small, which indicates its higher efficiency in detecting and classifying aircraft. The obtained results confirm the prospects of applying machine learning methods in the field of monitoring and security systems, which reliably detect and track unmanned aerial vehicles in various conditions. The high performance of the trained models demonstrates their effectiveness in real-world operating conditions, making them a valuable tool for solving important control and supervision tasks
Publisher
National University of Life and Environmental Sciences of Ukraine
Reference40 articles.
1. [1] Alam, K.M.R., Siddique, N., & Adeli, H. (2020). A dynamic ensemble learning algorithm for neural networks. Neural Computing and Applications, 32(12), 8675-8690. doi: 10.1007/s00521-019-04359-7.
2. [2] Bejani, M.M., & Ghatee, M. (2020). Theory of adaptive SVD regularization for deep neural networks. Neural Networks, 128, 33-46. doi: 10.1016/j.neunet.2020.04.021.
3. [3] Bhargavi, K. (2021). Deep learning architectures and tools: A comprehensive survey. In Deep learning applications and intelligent decision making in engineering (pp. 55-75). Hershey: IGI Global. doi: 10.4018/978-1-7998-2108-3.ch002.
4. [4] Bubryak, O.S., & Potapova, K.R. (2023). System of transportation vehicles detection on the basis of linear method SVM and neural network YOLO. In Proceedings of IX International Scientific and Practical Conference “Science and Innovation of Modern World” (pp. 206-210). London: Cognum Publishing House.
5. [5] Chen, L., Li, S., Bai, Q., Yang, J., Jiang, S., & Miao, Y. (2021). Review of image classification algorithms based on convolutional neural networks. Remote Sensing, 13(22), article number 4712. doi: 10.3390/rs13224712.