Improving SLAM in pipe networks by leveraging cylindrical regularity

Monocular visual Simultaneous Localisation and Mapping algorithms estimate map points and frame poses simultaneously based on video data. The estimated map point locations do not contain any structural information. Due to the measurement noise, the estimated trajectory is slightly different from the ground truth. This paper improves the estimation accuracy of trajectory in a pipe network by leveraging structural regularity. An optimisation-based method is used to detect a cylinder among map points in the SLAM back-end. When the cylinder is detected, the system enforces cylindrical regularity to the points from the cylindrical pipe surface, which is named cylindrical points. The estimated trajectory and map points will benefit from this structural information. This method is verified and evaluated on both synthetic data and real-world pipe video datasets.