Anomaly detection in wastewater pipeline videos using self-attention

Data related to urban infrastructures often come from multiple sources and exist in a wide variety of formats, such as Geographic Information Systems (GIS), textual information, numerical databases, images, or videos, which can make their processing, querying, and analysis complex. This work falls within this context and aims to propose new approaches for the management of heterogeneous data in stormwater and wastewater networks.

More specifically, we focus on video data, particularly Closed-Circuit Television (CCTV) inspection videos of sewer pipelines. These videos are essential for the management and maintenance of urban networks. On the one hand, they enable the identification of anomalies that may affect the integrity of pipelines, such as blockages or structural degradation. On the other hand, they provide key information on the structural properties of pipelines and networks, including pipe diameter and the direction of wastewater flow.

We propose a classification algorithm for wastewater inspection videos aimed at detecting major anomalies in CCTV inspection sequences of sewer networks, with a particular emphasis on identifying variations in pipe diameter, internal cracks, chemical corrosion, and the presence of turbid water within the pipelines. This task is crucial for predictive maintenance and hydraulic modeling of sewer systems. Information related to the identification of variations in pipe diameter can also be leveraged to enrich and complete missing pipe diameter attributes in Geographic Information Systems.

Our approach is based on the Video Vision Transformer (ViViT) and TimeSformer architectures, which effectively capture both spatial and temporal relationships in video data. We also describe various methodologies for generating training datasets from a subset of manually annotated images. Experimental results obtained on real-world CCTV sewer inspection videos provided by Montpellier Méditerranée Métropole demonstrate promising performance in anomaly detection.