Stereo Vision SLAM for Real-time Underwater Dense 3D Mapping

Visual Simultaneous Localization and Mapping (V- SLAM) still faces challenges in underwater environments, as existing methods often fail to deliver robust localization and globally consistent mapping due to uneven lighting, image degradation, and repetitive textures. We propose a novel stereo camera-based V-SLAM approach that enables real-time dense 3D reconstruction with high global localization accuracy in underwater environments. Firstly, a multi-level optical flow module is incorporated to optimize the selection of key viewpoints and to ensure real-time performance. Then, we introduce a stereo matching module to generate real-scale depth maps based on the key viewpoints for building dense mapping. Finally, a geometric-consistent checking module is employed to reduce ghosting artifacts in scene reconstruction and enhance the realism of local details. Extensive experiments conducted on both public datasets and our self-collected underwater datasets have demonstrated that our method achieves a superior speed-accuracy trade-off compared with state-of-the-art dense 3D mapping methods. Our demo is available at https://youtu.be/9K19M7Icew8.