Using machine learning-based automation process to improve the productivity of CCTV inspections of municipal drainage systems

点击翻译按钮获取中文摘要

Municipal drainage systems are probably one the most fundamental infrastructures which contributed to making modern cities marvels of civilisation that improved hygiene and by the same token general public health and quality of life. In fact, research in public health has clearly correlated sanitation and personal hygiene to longevity and decreased mortality (especially in infants). However, according to worldwide statistics, the current condition of public infrastructure (including roads, bridges, drainage systems, etc.) is already in a dire state and the financial requirements to bring it back to an acceptable level of service through better maintenance, repair, rehabilitation and replacement are now ranging in trillions of dollars. This situation can worsen even further for cities that are still expanding horizontally since in this case new infrastructure will be added to the existing inventory and thus will need to be included when forecasting future budgets for maintenance, repair, etc. In the case of drainage systems, structural and operational damage can be caused by a variety of factors, including natural ageing of pipes resulting from chemical reaction with sewage gases, soil movements and residential and commercial activities which can lead to material deposits. As a result, municipalities have devised a series of activities falling into two categories: cleaning and inspection, aiming at keeping this infrastructure operational at all time. Although time consuming and costly with respect to other inspection methods, CCTV is the most important method used to assess the condition of sewer pipes. From a data analysis perspective, CCTV inspection, which constitutes the focus of this research, can generically be viewed as a two-step process: (i) data (i.e. video) collection, and (ii) data analysis (i.e. video assessment) in the course of which defects are detected and classified. The proposed research aims at extracting additional value from CCTV inspection videos by building an automatic framework from which a condition-based deterioration model will be developed and applied to quantify the consequences of false negatives and misclassifications. Since the methodology underlying this research is anchored in the area of computer vision and image processing, the knowledge gained from building the analysis framework associated with CCTV inspections of sewer pipes can be adapted to other maintenance situations using a similar inspection technology.