Résumé
The management of wastewater networks is essential to ensure optimal hydraulic modeling. However, data on wastewater systems, usually represented in Geographic Information Systems (GIS), are often incomplete. Other sources of information are thus required to complete them. Among the potential sources, pipe inspection videos-traditionally used for maintenance and detecting issues, such as structural degradation, offer valuable, underexploited information. This article goes in this direction and proposes to take advantage of these inspection videos to extract the structure of wastewater networks, represented here in the form of graphs. For this, we propose automatic detection mechanisms of the different annotations associated with the inspection videos. These annotations contain enough information to construct a graph of the wastewater networks.Regular expressions and text recognition tools, based on optical character recognition (OCR), are used to extract manholes identifiers, water flow direction in pipelines, as well as other useful information to determine the geographical positions of wastewater network objects. In addition to the topological reconstruction, effective management also requires assessing the condition of each pipeline segment within the graph structure. To this end, we employ deep learning models-specifically, the Swin Transformer and MobileNetV2, for detecting anomalies and assessing pipe integrity. Experiments conducted on real-world data validate the effectiveness of our approach, underscoring its potential to improve both wastewater network management and hydraulic modeling accuracy.