Lightweight semantic 3D mapping in sewer pipes leveraging cylindrical geometry

In this paper, we present a lightweight method for 3D semantic mapping in sewer pipes from Closed-Circuit Television (CCTV), leveraging the cylindrical pipe geometry. The use of a cylindrical prior overcomes the problems with monocular cameras related to depth ambiguity, scale drift, and the need to perform computationally intensive 3D reconstruction for mapping. An additional contribution is that we use weakly supervised semantic segmentation for landmark detection, which avoids the need for pixel-level labelled datasets. The complete system uses separate extended Kalman filters (EKFs) for landmark tracking in the 2D image plane, which avoids problems due to low parallax for distant objects in pipes. Results on real-world, public sewer inspection data demonstrate that the perception module achieves an AUROC above 0.93, while the mapping backend operates in real-time at 27.63 FPS with centimetre-scale longitudinal accuracy (0.12 m).