Proud-SLAM: Neural Point-based Hybrid RGBD Monocular Dense SLAM

Abstract

Neural Radiance Fields (NeRF) have shown significant potential in encoding complex 3D geometries and visual properties, making them a promising alternative to traditional dense SLAM systems. We propose Proud-SLAM, a dense SLAM framework that merges a neural point-based hybrid scene representation. Our key innovation is the HPoint method, which integrates color features into the point cloud, preserving real-world fidelity and enhancing scene rendering. Additionally, we design a hashV structure to efficiently manage and expand point clouds incrementally. Experiments on Replica and Scan-Net datasets demonstrate that Proud-SLAM excels in tracking precision, reconstruction accuracy, and scene rendering quality.