A deep learning-based framework for an automated defect detection system for sewer pipes

Xianfei Yin*, Yuan Chen, Ahmed Bouferguene, Hamid Zaman, Mohamed Al-Hussein, Luke Kurach

*Corresponding author for this work

Research output: Contribution to journalArticleAcademicpeer-review

132 Citations (Scopus)


The municipal drainage system is a key component of every modern city's infrastructure. However, as the drainage system ages its pipes gradually deteriorate at rates that vary based on the conditions of utilisation (i.e., intrinsic conditions) and other extrinsic factors such as the presence of trees with deep roots or the traffic load above the sewer lines, which collectively can impact the structural integrity of the pipes. As a result, regular monitoring of the drainage system is extremely important since replacement is not only costly, but, more importantly, can disturb the daily routines of citizens. In this respect, closed-circuit television (CCTV) inspection has been widely accepted as an effective inspection technology for buried infrastructure. Since sewer pipes can run for thousands of kilometers underground, cities collect massive amounts of CCTV video footage, the assessment of which is time-consuming and may require a large team of trained technologists. A framework is proposed to realize the development of a real-time automated defect detection system that takes advantage of a deep-learning algorithm. The framework focuses on streamlining the information and data flow, proposing patterns of input and output data processing. With the development of deep learning techniques, a state-of-the-art convolutional neural network (CNN) based object detector, namely YOLOv3 network, has been employed in this research. This algorithm is known to be very efficient in the field of object detection from the perspective of processing speed and accuracy. The model used in this research has been trained with a data set of 4056 samples that contains six types of defects (i.e., broken, hole, deposits, crack, fracture, and root) and one type of construction feature (tap). The performance of the model is validated with a mean average precision (mAP) of 85.37%. The proposed output of the system includes labeled CCTV videos, frames that contain defects, and associated defect information. The labeled video can serve as the benchmark for assessment technologists while the multiple output frames provide an overview of the condition of the sewer pipe.

Original languageEnglish
Article number102967
JournalAutomation in construction
Early online date29 Oct 2019
Publication statusPublished - Jan 2020
Externally publishedYes


  • Automated detection
  • CCTV
  • Deep learning
  • Defects
  • Sewer pipe

Cite this