Abstract
Routine CCTV surveys are vital to the effective maintenance of wastewater networks, but their time-consuming nature makes them very expensive. We present a methodology capable of automatically detecting faults within recorded CCTV footage, aiming to improve surveying efficiency. The procedure calculates a feature descriptor for each video frame, before using a machine learning classifier to predict the contents of individual frames. The sequence of predictions is then smoothed using a Hidden Markov Model and order oblivious filtering, incorporating information from the entire sequence of frames. This technique has been demonstrated on footage collected by the Wessex Water, achieving a detection accuracy of over 80% on still images. Furthermore, temporal smoothing on continuous CCTV footage improved false negative rate by more than 20%, to achieve an accuracy of 80%. This last step enables the method to compete with the performance of trained technicians, showing promise for application in industry.
Original language | English |
---|---|
Pages (from-to) | 64-71 |
Number of pages | 8 |
Journal | Automation in Construction |
Volume | 95 |
DOIs | |
Publication status | Published - 1 Nov 2018 |
Externally published | Yes |
Keywords
- Fault detection
- GIST
- Hidden Markov model
- Random forest
- Sewer pipe
- Support vector machine