Naslov
Unexploded Ordnance Detection on UAV Thermal Images by using YOLOv7

Autori
Bajić, Milan ; Potočnik, Božidar

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), stručni

Skup
ROSUS 2023 - Računalniška obdelava slik in njena uporaba v Sloveniji 2023: Zbornik 17. strokovne konference

Mjesto i datum
Maribor, Slovenija, 23.03.2023

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Podatak o recenziji nije dostupan

Ključne riječi
unmanned aerial vehicle ; unexploded ordnance ; thermal imaging ; UXOTi_NPA dataset ; convolutional neural networks ; deep learning ; YOLO

Sažetak
A few promising solutions for thermal imaging Unexploded Ordnance (UXO) detection were proposed after the start of the military conflict in Ukraine in 2014. Our research focuses on improving the accuracy of UXO detection in thermal images. The current state-of-the-art UXO detection method is based on the YOLOv5 Convolutional Neural Network (CNN). We accessed the effectiveness of UXO detection by using the state-of-the-art object detector YOLOv7 in this article. Two YOLOv7 models were re-implemented, fine-tuned using a grid-search approach and trained on a UXOTi_NPA public dataset of 720x480 pixel thermal images. The results showed that the models were able to identify UXOs from 11 different classes with more than 90% probability and a Mean Average Precision (mAP) of 86.8% to 89.7%, depending on the model's complexity. The metrics are just slightly behind the YOLOv5 results. Such CNN, thus, enables accurate automatic UXO detection, which is crucial to address one of the least explored and life-threatening problems worldwide.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo

POVEZANOST RADA