Abstract Foreign object detection (FOD) is a crucial concern in various industries, including aerospace, automotive, and manufacturing, where undetected foreign objects can cause severe damage and safety risks. Traditional FOD detection systems rely on visual inspection, radar, X-ray imaging, or acoustic sensing, but face challenges in complex and dynamic environments. In this work, we propose a novel methodology that utilizes neural radiance fields (NeRF) as a comparative baseline for foreign object detection in industrial settings. Our approach consists of two main phases: First, a photorealistic 3D reconstruction of the environment is generated using NeRF when no foreign objects are present. Then, in the detection phase, new images of the environment are captured, and their viewpoints are matched with the NeRF-rendered scene to identify discrepancies using a deep learning-based comparative model. The system utilizes a modified SNUNet-CD architecture, enhancing attention mechanisms for improved accuracy in detecting semantic changes. Experimental validation in a controlled environment demonstrates high detection accuracy, successfully identifying foreign objects of varying sizes, materials, and placements.
Lopez et al. (2026) studied this question.