Temporal defect point localization in pipe CCTV Videos with Transformers

During the inspection and maintenance of underground pipe systems, technicians often spend considerable time searching for subtle defects in inspection videos captured under varying pipe conditions using Closed-Circuit Television (CCTV). The lack of feature extractors tailored for pipe images, combined with the complexity of pipe CCTV videos, poses substantial challenges to the performance and applicability of conventional frame-by-frame, image-based localization algorithms. To address these challenges, this paper introduces PipeTR, a transformer-driven, end-to-end network, offering enhanced insights into pipe CCTV video analysis by shifting from a frame-based to a video-based approach. The development of PipeTR aims to assist technicians by automating the most time-consuming step of the assessment, thereby improving both efficiency and accuracy. Experiments demonstrate that PipeTR outperforms other image-based, frame-by-frame analysis methods on real-world CCTV pipe inspection video datasets, achieving an average F1 score of 43.04%, which is a 5.35% improvement over the current state-of-the-art methods.