A Deep Dive into Computer Vision Aided Sewer Inspections

The sewerage infrastructure is one of the critical infrastructures of modern society, which most people rarely consider. However, due to its immense size regular inspection of the sewer pipes is impossible. This thesis focuses on using Computer Vision to automate sewer inspections through two considered modalities: images and point clouds. Computer vision aided sewer inspections has been researched for over three decades but has yet to be widely adopted by professional inspectors.

In this thesis, the fundamental historic trends and hindrances were investigated, covering the algorithmic trends and the lack of public code and datasets as well as no common evaluation protocols. These hindrances were broken down by the release of the first publicly available image and point cloud sewer defect classification datasets, the introduction of domain influenced evaluation metrics, and open-sourcing the developed code. This has consequently made the automated sewer inspection domain far more accessible. Finally, two novel graph-based computer vision algorithms were developed for automating parts of the sewer inspection process leading to significant improvements over prior methods.