| dc.description.abstract | Sewers across the globe must be regularly inspected to ensure their smooth running and effective maintenance. Furthermore, surveys are often performed reactively, often diagnosing suspected faults within a network. Almost all surveys in the UK and abroad are performed using closed circuit television (CCTV) cameras where pipes are too small for manual inspection. As such vast quantities of footage are recorded by surveying teams on a daily basis. This footage is currently analysed manually, requiring a trained engineer to watch through its entirety, annotating potential faults. This thesis examines methods of improving this labelling process, implementing various machine learning and image processing techniques to automate this procedure. The thesis presents two distinct methodologies: the first for the detection of faults, using only raw CCTV footage, whilst the second identifies the type of a detected fault according to the Manual of Sewer Condition Classification. The fault detection methodology identifies the presence of a fault within a CCTV image. The methodology calculates a GIST feature descriptor for each video frame, before utilising a Random Forest classifier, to predict the presence of a fault. The basic methodology was further refined with the inclusion of smoothing, to eliminate isolated inconstancies, and stacking to intuitively combine the results of multiple machine learning classifiers. The final methodology achieved a detection accuracy of 86% on unseen real-life data from the UK. The fault classification methodology identifies the fault type in images, where faults have been previously detected using the above technique. The tool again calculates a frame’s GIST descriptor before applying multiple Random Forest classifiers in a ‘1 vs all’ architecture to predict the type of a given fault. This architecture allowed for comparative classifications and later enabled the identification of multiple faults within a single frame. The methodology achieved a peak accuracy of 74% when classifying faults well represented by the dataset (at least 100 examples). Furthermore, when including multi-label functionality, the tool achieved an accuracy of 67% across all fault types. Both methods have been developed to be holistic and practical, utilising only industry standard CCTV footage and generalising well across all types of sewer system (size, shape and material). Furthermore, as both methodologies rely on the same feature descriptor, they integrate well to form a methodology that could be applied in real time. As such the thesis also explores the practical implications of creating a detection support tool capable of integration with current working practices. In combination both methodologies and their additions present a unique contribution to the field of automated sewer surveying. Achieving competitive accuracies with a streamlined methodology, the technology shows promise for future application in industry, greatly increasing the speed, accuracy and consistency of CCTV sewer surveys. | en_GB |
