Evaluation of the Robustness-Runtime Efficiency Trade-Off of Edge AI Models in UXO Localisation and Classification
Gheorghe-Marian Craioveanu, Grigore Stâmâtescu · 2025
Real time localisation and classification of Unexploded Ordnance (UXO) can significantly benefit from advanced new model compression and quantization techniques towards embedded deployment on resource constrained fixed or mobile hardware platforms. This can extend the applicability, usefulness and adoption by first responders of such methods in real-world scenarios with significant social and environmental benefits. The proposed methodology considers the emergence of multiple frameworks and tools that have now become available to automate the comparative assessment of state-of-the-art image classification edge AI model. As main results, we present a quantitative evaluation of the robustness-runtime efficiency trade-off for representative CNN-based vision model and a parametrization discussion on a reference public UXO dataset. The approach is validated through deployment and experiments using a reference embedded GPU development board i.e. the Nvidia Xavier NX.