Under Water SLAM with Laser-light sectioning method using ST-GAT

Introduction

The demand for high-density and accurate 3D shape measurement of underwater scenes is rapidly growing for critical applications such as infrastructure inspection and marine resource exploration. However, the underwater medium poses unique challenges, such as light attenuation and scattering, which thwart conventional 3D sensing techniques. While multi-line laser systems offer a practical solution, they suffer from laser line fragmentation due to occlusion and scattering, leading to the failure of traditional ID assignment methods based on epipolar geometry.

In this work, we reformulate the laser ID assignment problem as a graph-based sequence labeling task and propose a novel two-stage hierarchical framework using Spatio-Temporal Graph Attention Networks (ST-GAT).
Figure 1: Comparison of reconstruction results on a challenging underwater sequence with fragmented laser lines. Our method robustly reconstructs a complete and dense model compared to the sparse baseline.

Methodology

Our core contribution is a robust GNN-based module for multi-line laser ID assignment. The proposed framework operates in two hierarchical stages to ensure both local topological correctness and global temporal consistency:
  • Stage 1: Node-Level Spatio-Temporal Prediction. We construct a spatio-temporal graph from laser endpoints and intersections. A custom ST-GAT with a unified attention mechanism captures topological cues to predict initial IDs, handling local fragmentation effectively.
  • Stage 2: Segment-Level Trajectory Optimization. We elevate reasoning to the segment level. By constructing segment trajectories and applying trajectory-constrained Viterbi decoding, we enforce physical consistency, ensuring that a laser line maintains a single identity throughout the sequence.
This approach eliminates the reliance on complete epipolar geometry, making it highly robust in turbid underwater environments.
Figure 3: (A) The overview of our underwater 3D reconstruction system. (B) The proposed ST-GAT method, illustrating the Stage 1 node-level prediction and Stage 2 segment-level trajectory optimization.

Experimental Results

We conducted comprehensive experiments in real underwater environments, including a clear pool and a scattering water tank. We compared our method against state-of-the-art passive methods (COLMAP, VGGT, 2DGS, WaterSplatting) and traditional active structured-light methods.

As shown in the qualitative results below, our method demonstrates superior reconstruction completeness and temporal stability, especially in challenging environments where passive methods suffer from noise and traditional active methods fail due to data loss. Quantitative evaluations confirm that our approach achieves the best balance of accuracy and coverage.
Figure 7: Qualitative comparison in a challenging scattering environment. Our method (g) yields the most complete and structurally accurate reconstruction compared to other SOTA baselines.

Resources
    Dataset: Zenodo (The underwater video dataset)

Publications
Computer Vision and Graphics Laboratory