Using Automatic Anomaly Detection to Identify Faults in Sewers

CCTV surveys are a common practice within water companies worldwide, and necessary to ensure the effective maintenance of all sewer pipes. Currently surveys are time consuming, requiring a trained technician to watch and label hours of video footage. This paper presents an automated anomaly detection approach for the detection of sewer faults using only raw CCTV sequences. A One-Class Support Vector Machine (OCSVM), is trained using images of normal pipe, and highlights any abnormalities or faults within a sewer survey for further investigation. This technique is validated and demonstrated on a comprehensive dataset of ~8000 still CCTV images, achieving an accuracy of 75% and an accuracy of 85% when applied to a video sequence. Overall the OCSVM proves itself to be an effective anomaly detection technique.