Automated Detection of Faults in Wastewater Pipes from CCTV Footage by Using Random Forests

Abstract

Sewer systems require regular inspection in order to ensure their satisfactory condition. As most sewer networks consist of pipes too small for engineers to traverse, CCTV footage is used to record the interior of these pipes. This footage is manually analysed by qualified engineers, to determine the condition of the pipe and the presence of any faults. We propose a methodology, which automatically detects faults within the CCTV footage. This has the potential to dramatically reduce the time required to process the large volume of CCTV footage produced during a survey. The proposed methodology first characterises localised regions of each video frame using multiscale GIST features. Extremely randomised trees are then used to learn a classifier that distinguishes between frames showing a fault and normal frames. The technique is tested on 670 video segments from real sewer inspections of a variety of pipes, supplied by Wessex Water. Detection performance is assessed by plotting receiver operating characteristics and quantifying the area under the curve. Preliminary results indicate high detection accuracy of 88% and an area under the ROC curve of 96%. The machine learning used reduces the footage to a selection of frames containing faults, which can be quickly identified (whether by an engineer or another piece of software), showing promise for use in industrial wastewater network surveys.