Towards an automated condition assessment framework of underground sewer pipes based on closed-circuit television (CCTV) images. (April 2021)

• Record Type: Journal Article
• Authors: Wang, Mingzhu; Luo, Han; Cheng, Jack C.P.
• Is Part Of: Tunnelling and underground space technology. Volume 110(2021)
• Publication Date: 2021-04
• Subjects: Condition assessment -- Computer vision -- Defect detection -- Defect segmentation -- Severity assessment -- Sewer pipe; Tunneling -- Periodicals; Underground construction -- Periodicals; Tunnels -- Periodicals; Underground areas -- Periodicals; 624.193
• DOI: 10.1016/j.tust.2021.103840
• Languages: English
• ISSNs: 0886-7798

Abstract

Highlights: A generic framework for automated defect severity and pipe condition assessment. Sewer defect severity is evaluated by computer vision approaches. Pipe cross section area is measured accurately based on joint detection and fitting. The framework generates condition assessment results consistent with inspectors. Abstract: There is a growing trend of using computer vision techniques for interpreting Closed-Circuit Television (CCTV) inspection videos of sewer pipes. Previous studies mainly focus on detecting defect types and locations in CCTV images, yet limited systematic approaches are available for automatically evaluating defect severity and sewer condition with references to existing standards. This study proposes a framework for evaluating defect severity and sewer condition automatically from CCTV images using computer vision methods, which includes (1) required information definition for sewer condition assessment, (2) pipe joint detection and fitting by image processing techniques to obtain cross section area, (3) defect detection and segmentation to obtain defect type, location and area, and (4) evaluation of defect severity and sewer condition. Particularly, three deep learning -based defect detection models are developed, among which the model based on Faster R-CNN (regional convolution neural network) outperforms others with higher accuracy and is used for detecting defect type and location in the image. Meanwhile, an innovative semantic segmentation model is applied for segmenting defects to extract defect area in the image. In the validation, our framework performs well in defect detection with an average precision, recall and F1 of 88.99%, 87.96%, and 88.21% respectively. More importantly, the framework evaluates Operation and Maintenance (O&M) defects more accurately by precise calculation and generates the overall condition gradings that are mostly consistent with professional inspectors, only with an average deviation of 3.06%. Our framework can assist the review of inspection videos and lays the basis for fully automated sewer assessment and maintenance planning in the future. Without constraints on the assessment codes and computer vision methods, the framework is adaptable to evaluating sewer condition in different regions and can achieve better performance by integrating with cutting-edge vision techniques.