Assessing the feasibility of monocular visual simultaneous localization and mapping for live sewer pipes: a field robotics study

Sewer pipes are important to inspect for damage and blockages. Mobile robots with cameras are a natural choice for inspecting sewers, and indeed CCTV inspection using tethered mobile platforms is a well-established commercial approach. It therefore makes sense to also explore the use of camera data for localising defects for targeting subsequent repair. Visual odometry (VO) methods have been researched for robot localisation in pipes but the full visual simultaneous localisation and mapping (vSLAM) problem has received little attention. Whilst VO focuses on estimating the current pose of the robot, vSLAM focuses on building a map, as well as pose estimation, which should increase accuracy and robustness - both important for the future use of autonomous robots in sewer inspection. In particular, it is not known if one crucial element of vSLAM - loop closing using appearance-recognition methods - works effectively in sewer pipes due to problems of perceptual aliasing - where the high degree of visual similarity in image frames can lead to incorrect loop closures causing the vSLAM system to fail. The aim of this paper is to assess the feasibility of vSLAM for sewer pipes using real world data. The results demonstrate that whilst perceptual aliasing is a problem, appearance-recognition using bag-of-words methods can be used effectively. Demonstrating for the first time that vSLAM systems are potentially useful for sewer pipe environments.