Deep Learning–Based Automated Detection of Sewer Defects in CCTV Videos

Automated interpretation of closed-circuit television (CCTV) inspection videos could improve the speed and consistency of sewer condition assessment. Previous approaches focus on defect classification, with less emphasis on defect localization. Furthermore, previous approaches used pre-engineered features for image classification, leading to low generalization capabilities. This paper presents a deep learning–based framework for the classification and localization of sewer defects. Three state-of-the-art models—single-shot detector (SSD), you only look once (YOLO), and faster region-based convolutional neural network (Faster R-CNN)—are evaluated for speed and precision in detecting sewer defects. Three thousand eight hundred annotated images of defects were used to train and test the models. To demonstrate the viability of real-time automated defect detection, a prototype system was developed for detecting root intrusions and deposits, and evaluated on inspection videos televising 335 m of sewer laterals. The prototype system detected 51 out of 56 instances of defects and generated four false positives. Future research aims to incorporate postprocessing and data fusion to improve the speed and accuracy of the prototype.