Exploiting Video Inspection Data in Wastewater Networks

In this presentation, we introduce an algorithm for extracting the structure of a wastewater network from a set of sewer inspection videos. This structure is represented as a directed graph of the pipes, automatically constructed from annotations present in the sewer videos. These annotations contain summary information about the inspection process. They include manhole identifiers, direction of inspection, direction of wastewater flow, distance travelled, date of inspection, name of the street where the pipe is located, etc. This graph, where the nodes represent manholes and the directed arcs represent pipes and wastewater flow, will provide valuable data to complement and compare with existing Geographic Information Systems. However, its construction is challenging due to the variable visibility of text in inspection videos, influenced by background brightness and irregular annotation positioning. By leveraging recurring annotations across multiple frames and using fusion strategies as well as regular expressions, we achieve reliable detection of key information such as street names and manhole identifiers, confirmed by experimental results on real wastewater inspection videos.