Oklahoma State University, and researchers led by Jasper Baur and Gabriel Steinberg, co-founders of the Demining Research Community, a nonprofit organization bridging academic research and humanitarian demining efforts have been in Oklahoma for two weeks, setting up grids of mines and munitions to train a drone-based, machine-learning-powered detection system.
Leveraging what we believe to be the most robust and extensive drone-based imagery dataset of surface munitions, we are training a convolutional neural network (CNN) to detect and categorize over 50 different types of landmines and UXO. Categorizations will include projectiles, anti-personnel landmines, anti-vehicle landmines, 40mm grenades and cluster munitions.
This method is designed to work for surface munitions on any environment with munitions ranging from fully visible to heavily obscured. Initial tests yield over 70% accuracy averaged across all objects; some larger targets yield a near 100% detection rate while camouflaged or easily obscured targets often have a lower accuracy.
Sources: YouTube; Demining Research Community