Unsupervised Stationary Flow Filter in CCTV Pipeline Inspection Videos

CCTV Pipeline Inspection is considered a critical component of sewage utility operations and maintenance, yet the reliance on human inspection within the water industry has been observed to become increasingly inefficient. An extensive inspection video archive exists without adequate exploratory tools. Manual review processes often entail extensive time and effort, often encountering stationary frames when operators pause at pipeline segments. Acknowledging these inefficiencies, the focus of this study has been on the development of an algorithm aimed at augmenting human video review processes by utilizing stationary characteristics.A two-stage workflow is employed: firstly, optical flow and perceptual similarity are computed as features for consecutive frames within the video. Subsequently, the Stationary Flow Filteralgorithm is introduced which is a time series signal processing filter designed to identify stationary frames. This technique demonstrates its effectiveness through the application of both scalar and quadratic curve fits to the filtered stationary flow magnitude computed over a video. An unsupervised approach is utilized to classify stationary frames, with a flow magnitude threshold estimated using Fourier Analysis. Samples from the Video Pipe dataset were labeled for benchmarks in stationary frame classification. Remarkably, an improved F1 score due to over 2x improvement in recall compared to traditional supervised machine-learning methods is showcased by our unsupervised model. The implementation of the stationary flow filter method holds promise in emphasizing specific frames or segments within inspection videos, significantly streamlining the coding and review processes.