Towards Automated As‐Built Documentation of Underground Utilities With Photogrammetry and Deep Learning

Urbanization has profoundly transformed city landscapes, leading to the densification of underground infrastructures. In response to this challenge, a regulation was introduced in France in 2012 aimed at improving the mapping of utility networks, which had become outdated and imprecise in many areas. This regulation promotes the adoption of advanced technologies for the localization and mapping of these infrastructures. It is within this context that the use of photogrammetry as a novel source of data for the as‐built documentation of underground utilities has gained interest. Unlike traditional methods, photogrammetry offers rapid deployment, easy data updates, and efficient coverage of large open trench areas. Moreover, the data collected through photogrammetry can be integrated into geographic information systems (GIS), enabling more detailed visualization and analysis of subsurface environments. In this paper, we propose to further enhance photogrammetry models by incorporating deep learning‐based object recognition, allowing attribute‐rich information to be overlaid on the point cloud. This combination of photogrammetry with artificial intelligence (AI) enriches the resulting model with contextual insights, making it more informative and useful for complex analysis and decision‐making. Achieving this vision requires the development of a deep learning ecosystem, including the creation of a specialized dataset for pipe identification in images and the selection of an appropriate neural network architecture for instance segmentation tasks. Based on this ecosystem, a vectorization algorithm will be in charge to create the corresponding as‐built plan of the covered trench.