Deep learning-based detection of surface and buried landmines

Edwards, Thomas; Nibouche, Mokhtar; Withey, Daniel

Abstract

Landmines are explosive weapons that can lie dormant for years and cause disastrous effects on the local populations, who often must learn to live with the threat of landmines. Clearing, or even simply identifying landmines, is dangerous and challenging work, which often requires expensive equipment and highly skilled operators. Fast, reliable and accessible methods for identifying landmines are required so that it becomes easier and quicker to identify them, allowing communities to live free from risk posed by landmines. This paper investigates how novel systems using thermography and Machine Learning (ML) can be used to detect surface and buried landmines. It also highlights significant environmental challenges with obtaining images and the considerations that should be made to mitigate the risk of obtaining images that can be difficult to use. A ML model, utilising a Convolutional Neural Network (CNN) and YOLOv8, is tested with a realistic and challenging dataset to assess its capability in being used to automatically locate landmines in Infrared (IR) images. Impressive results are achieved, with the CNN predicting the correct outcome 92.31% of the time, demonstrating the potential of this novel solution.

• Presentation Conference Type: Conference Paper (published)
• Conference Name: 2024 30th International Conference on Mechatronics and Machine Vision in Practice (M2VIP)
• Start Date: Oct 3, 2024
• End Date: Oct 5, 2024
• Online Publication Date: Nov 12, 2024
• Publication Date: Nov 12, 2024
• Deposit Date: Jan 9, 2025
• Publicly Available Date: Jan 10, 2025
• Publisher: Institute of Electrical and Electronics Engineers (IEEE)
• Peer Reviewed: Peer Reviewed
• Book Title: 2024 30th International Conference on Mechatronics and Machine Vision in Practice (M2VIP)
• ISBN: 9798350391923
• DOI: https://doi.org/10.1109/m2vip62491.2024.10746152
• Public URL: https://uwe-repository.worktribe.com/output/13461782