Detection of All Dynamic Objects in Neural SLAM with Local Voxel Map

Recent Visual SLAM research focus on scenes that are static, where minimal changes are assumed. However, these scenarios do not consistently align with the dynamic nature frequently encountered in real-world environments. We present LVM-SLAM, a dense neural visual SLAM, which employs geometry constraints to build the local voxel map and detect all dynamic objects. The implemented local voxel map is low-cost and efficient for filtering out dynamic points. Our system can accurately track in highly dynamic situations with numerous moving targets, and remove dynamic targets during the 3D scene reconstruction process. We evaluate our system on TUM RGB-D and Bonn Dynamic RGB-D datasets. Experimental results reveal that our system not only improves the tracking accuracy, but also enhances the real-time performance without the necessity for introducing a segmentation network.