UAV Thermal Imaging for Unexploded Ordnance Detection by Using Deep Learning

A few promising solutions for thermal imaging Unexploded Ordnance (UXO) detection were proposed after the start of military conflict in Ukraine 2014. At same time, most landmine clearance protocols and practices are based on old, 20th-century technologies. More than 60 countries worldwide still affected by explosive remnants war, new areas contaminated almost every day. To date, no automated exist surface UXO using imaging. One reasons is also that there publicly available data. This research bridges both gaps introducing an method, publishing During a project Bosnia Herzegovina 2019, organisation, Norwegian People’s Aid, collected data about unexploded ordnances made them this research. Thermal images with size 720 × 480 pixels Unmanned Aerial Vehicle at height 3 m, thus achieving very small Ground Sampling Distance (GSD). goals our was to verify if war remnants’ accuracy could be improved further Convolutional Neural Networks (CNN). We have experimented various existing modern CNN architectures object identification, whereat YOLOv5 model selected as retraining. An eleven-class problem solved primarily study. Our annotated semi-manually. Five versions model, fine-tuned grid-search, trained end-to-end randomly 640 training 80 validation from dataset. The models verified remaining 88 Objects each eleven classes identified more 90% probability, Mean Average Precision (mAP) 0.5 threshold 99.5%, mAP thresholds 0.95 87.0% up 90.5%, depending model’s complexity. results comparable state-of-the-art, these methods been tested other similar datasets images. study one field Automated images, first solves identifying class objects. On hand, relatively GSD will enable stimulate development algorithms, where method can serve baseline. Only really accurate automatic help solve least explored life-threatening problems.