A Vision-based Control Framework for Real-time Autonomous UUV Operations
This work addresses the challenge of achieving robust autonomous navigation for unmanned underwater vehicles (UUVs) in dynamic, low-visibility underwater environments by proposing an end-to-end vision-driven control framework. For the first time, this approach integrates visual perception, real-time 3D mapping, and dual-mode local–global localization within a unified architecture. By fusing vision-based SLAM, deep learning–based feature extraction, and multimodal pose estimation, the method significantly enhances system robustness under complex disturbances while maintaining real-time performance. Experimental results demonstrate that the proposed system efficiently constructs consistent 3D maps on both synthetic datasets and real-world UUV platforms, reliably enabling autonomous navigation and critical mission deployment.