Anomaly-Aware Change Detection for Oil Refinery Inspection Using a Mobile Robot with Viewpoint-Aligned Novel View Synthesis

Stable operation of oil refineries is essential for ensuring a continuous supply of petroleum products, which underpin the foundation of modern society. However, various anomalies such as leaks can occur even during regular op- eration, making periodic inspection indispensable. To reduce the burden on human operators, automated visual inspection using mobile robots has been attracting increasing attention as a promising alternative to manual inspections. One promising ap- proach, which we refer to as anomaly-aware change detection, is to compare videos captured during past and current inspections to identify scene changes specifically caused by anomalies. However, the robot cannot perfectly retrace its previous path, resulting in viewpoint misalignment between the two videos, which significantly degrades the performance of naive frame- wise comparison methods. To address this issue, we propose a novel inspection method that reconstructs a 3D model of the refinery from the past inspection video using Structure from Motion and 3D Gaussian Splatting, and then renders novel view images from the same viewpoints as those in the current inspection video. This allows us to obtain geometrically aligned image pairs, enabling anomaly-aware change detection that is robust to viewpoint misalignment. Field experiments conducted in an operational oil refinery achieved an F1-score of 0.806, significantly outperforming conventional methods and demonstrating the effectiveness of our approach.