Structured Contour Matching with Graph Convolutional Networks for Underwater Visual Positioning in Ocean Engineering

High-precision positioning for underwater vehicles like ROVs and AUVs is a significant challenge in GPS-denied environments, especially when cooperative targets are unavailable. This paper proposes a visual positioning method leveraging structural information from non-cooperative objects in marine engineering scenes. The framework enhances degraded underwater images to extract salient contours. A Siamese Neural Network mitigates matching ambiguity through initial target recognition. The core of this work is a Graph Convolutional Network that robustly matches incomplete and distorted contour features against a standard model. This graph-based approach captures the structure’s topological properties, making it resilient to occlusions and noise. The vehicle’s 6-DOF pose is then calculated from 2D-3D correspondences using a RANSAC-based PnP algorithm. Experiments in a controlled pool, validated by a high-precision total station, show strong performance, achieving a positioning RMSE of approximately 4.65 cm. These results validate the method as a robust solution for correcting inertial navigation drift and enhancing underwater autonomy.