Memory-Efficient 3D LiDAR Graph SLAM for Ballast Water Tank Inspection Robots Using Robust Hierarchical Bundle Adjustment and a Kaczmarz Backend

Autonomous inspection of ballast water tanks requires three-dimensional (3D) LiDAR-based simultaneous localization and mapping (SLAM) in Global Positioning System (GPS)-denied, geometrically repetitive interiors, where sensing, mapping, and control modules share a limited onboard memory budget. Graph SLAM backends that rely on sparse factorization can incur fill-in, increasing peak memory and limiting deployment on edge computers. The proposed architecture couples a robust hierarchical bundle adjustment frontend with a factorization-free Kaczmarz backend. The frontend combines residual-adaptive weighting, damped and bounded pose updates, soft fallback, local-map compression, and memory-aware keyframe control. The backend stores the whitened Jacobian in compressed sparse row (CSR) format and performs row-wise projections without explicitly forming the normal equations, a Cholesky factor, or a transpose cache. Evaluation was conducted on Norwegian University of Science and Technology (NTNU) Ballast Water Tank missions 1–3, containing 851, 1202, and 1084 LiDAR frames. Following robust local bundle adjustment and verified similarity alignment, translational root-mean-square errors were 0.080, 0.110, and 0.127 m, corresponding to 0.137%, 0.143%, and 0.122% of the reference path lengths; archived baseline ratios ranged from 0.281% to 0.372%. These results support a numerical architecture that combines frontend stabilization, row-wise optimization, and memory-aware policies for resource-constrained marine inspection robots.