Robust and Low-Memory Consumption Online 3D Reconstruction Based on VDB and Relocalization

Scene reconstruction holds significant importance in the field of computer vision. Traditional 3D reconstruction methods usually need to obtain the scene image poses for offline reconstruction. They tend to have substantial memory consumption when reconstructing large scenes. To address these challenges, we propose a robust and low-memory consumption online incremental scene reconstruction method by integrating Simultaneous Localization and Mapping (SLAM) and the Truncated Signed Distance Function (TSDF) reconstruction algorithm. We employ relocalization to accurately obtain real-time scene image pose information, reducing the impact of drift in SLAM process. We integrate semantic information of scene objects for relocalization, enhancing the robustness of both relocalization and reconstruction. Additionally, we utilize the Voxel Data Base (VDB) data structure to implement the volumetric mapping of the TSDF algorithm, thus reducing memory consumption during scene reconstruction. Due to its low hardware requirements, our method is suitable for deployment on mobile robots with limited computational resources. Experimental validation on the TUM datasets, compared to state-of-the-art method, demonstrates that our method reduces memory consumption and enhances mapping completeness and accuracy.