Semantic Segmentation of Cargo Tank Point Clouds Using Structural Priors

Nikhil Khedekar, Kostas Alexis

Keywords: Pointcloud Semantic Segmentation, Maritime Inspection

TL;DR: A geometry-driven pipeline that semantically segments accumulated LiDAR point clouds of mid-section cargo tanks on oil tankers.

Abstract: Efficient autonomous inspection of cargo tanks in maritime vessels requires semantic understanding of the tank's structural components to target corrosion-prone members. To this end, we present a geometry-driven pipeline for semantic segmentation of 3D point clouds acquired inside cargo tanks. The method uses the known rectangular prismatic structure of mid-section cargo tanks to estimate a tank aligned coordinate frame from surface normal statistics, fit hull surfaces and the hopper plate via guided RANSAC, and detect webframes and stringer decks through histogram-based peak detection followed by guided RANSAC plane fitting with density-based clustering. The approach determines the number of webframes and stringer decks without prior knowledge of their count or positions, and operates on noisy, accumulated LiDAR maps. We evaluate on point clouds from multiple crude oil tankers collected by a commercial aerial inspection robot, detecting all seven hull surfaces, five webframes, and four stringer decks, with webframe spacings matching CAD reference values to within 0.01m across all four tested tanks.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 21