Automated interpretation of closed-circuit television (CCTV) inspection videos has the potential to improve the speed, accuracy, and consistency of sewer pipeline condition assessment. Previous approaches have focused on defect classification in images, with little focus on defect localization (i.e., calculating location of defects relative to the pipe). Recent studies have shown that deep-learning based object detection models can be used to classify and localize operational defects, such as roots and deposits; however, the detection of structural defects, such as pipe fractures is challenging, given their fine silhouettes in images. This paper presents a two-step defect detection framework that uses a 5-layered convolutional neural network (CNN) for classification followed by the you-only-look-once (YOLO) model for detection of pipe fractures. The framework was trained using 1,800 images and yielded a 0.71 average precision (AP) score in detecting fractures, when tested on 300 images. The proposed framework can also achieve a significant reduction in time taken to process CCTV videos. Ongoing research aims to validate the framework on videos from a variety of pipes and extend the framework to defect additional defect categories.
展开▼
