Underground sewer pipe condition assessment based on convolutional neural networks

Abstract

Surveys for assessing the condition of sewer pipeline systems are mainly based on video surveillance or CCTV, which is a time-consuming process that relies heavily on human labor because an operator has to watch videos, looks for defects and decides the defect's type manually. Previous research required suitable handcrafted features that were inefficient in analyzing sewer pipeline condition, so a robust and efficient framework is crucial as it eliminates the time-consuming tasks and helps the operator access condition of sewer systems more efficiently. This study proposes a defect classification system on CCTV inspection videos based on convolutional neural networks (CNN). The dataset was manually constructed and validated by extracting the images from CCTV videos, and the images were labeled according to six predefined defects. The CNN model was fine-tuned before training, and trained on a total of 47,072 images (256 × 256 pixels). The highest recorded accuracy was at 96.33%. As a result, the presented framework will motivate the finding of a more robust model that automatically and precisely evaluates the condition of sewer pipeline systems using CCTV and encourages the integration of the proposed model in real applications.