Abstract: This paper proposes a monocular vision-driven real-time high-precision dense scene reconstruction algorithm for robots, based on deep dense monocular visual SLAM and rapid uncertainty propagation techniques to reconstruct 3D scenes from images.The algorithm achieves dense, accurate, and real-time 3D scene reconstruction while demonstrating robustness against extreme noise in depth estimation from monocular visual SLAM.Unlike traditional methods that rely on specialized depth filters or estimate depth uncertainty from RGB-D sensor models, this approach directly utilizes the information matrix from the underlying bundle adjustment problem in SLAM to generate probabilistic depth uncertainty.This depth uncertainty provides a critical signal for weighting depth maps during volumetric fusion.Our method produces more precise 3D meshes with significantly reduced artifacts.Experimental validation on the challenging Euroc dataset shows that compared to methods that directly fuse depths from monocular visual SLAM improves mapping accuracy by 85%.

Key words: monocular vision, dense reconstruction, SLAM, depth uncertainty, robotics