关键词: 视觉SLAM, 三维重建, 姿态估计, 地图构建

Abstract: To enable a Simultaneous Localization and Mapping (SLAM) system to adapt to the dynamic changes in smart manufacturing environments and continuously improve its accuracy and robustness,a novel end-to-end RGB-D SLAM system named NPF-SLAM was proposed.The system employed a feature-based deep neural tracker as the frontend and utilized a Neural Radiance Fields (NeRF)-based neural implicit map builder as the backend.The neural implicit map builder was trained in real-time,while the neural tracker was pre-trained on datasets,such as Euroc,which were collected in real-world smart manufacturing environments.The tracker was also fine-tuned during the training of the map builder.With this design,NPF-SLAM could learn to track cameras using scene-specific features,enabling lifelong learning within the SLAM system.Additionally,both the tracker and the map builder were trained in a self-supervised manner,without requiring ground truth for poses.Performance evaluations on the Replica,ScanNet and Euroc datasets showed that NPF-SLAM outperformed existing NeRF-based SLAM systems in both scene reconstruction and camera tracking,demonstrating its adaptability and state-of-the-art performance in complex smart manufacturing environments.

Key words: visual simultaneous localization and mapping, 3D reconstruction, pose estimation, map construction