Research And Application Of Anomaly Detection For Urban Complex Underground Sewer Pipes

Urban underground sewer pipes are necessary for urban operation.However,due to the inefficient maintenance of sewers,most of them now have problems such as sub-sidence and settlement.To ensure the safety of staff,after using robots downhole to shoot sewer videos,it will be screened out from these videos which contain sewer pipe defects.This method commonly used in the maintenance of current sewer pipes.Nev-ertheless,during the sewer video screening stage,a large number of staff with relevant knowledge are need to responsible for this.The working mode at videos screening stage is the main reason for the inefficient maintenance of the sewer pipe.On the other hands,deep learning based methods has breakthroughs in all kinds of fields in recent years.Compared with traditional machine vision algorithms,these methods based on deep learning generally have the advantages of higher accuracy and better generalization performance.Thus,how to use deep learning methods to achieve efficient sewer video anomaly detection has become the core research content of this article.The following is the research content and several contributions of this article:1.We collected and annotated a large urban underground sewer pipes dataset;2.We propose a two-stream feature fusion neural network for urban underground sewer pipes anomaly detection;3.We successfully explored a multi-tasking mode to improve anomaly detection performance;4.In order to correct recognition of weak anomaly videos,we introduce a data preprocessing method named "Binary Frames Blending".We demonstrate our experiments on urban underground sewer pipes datasets achieve the outperformance in comparison with many anomaly detection state-of-the-art meth-ods.Our proposed method achieves 85%precisions when the abnormal recall reaches 99%.